[XLA] Use generic fusion's multi-output fusion support.

The main benefit here is a 6% speedup on MLPERF ncf.

BEGIN_PUBLIC
[XLA] Fix multioutput fusion support in normal instruction fusion.
END_PUBLIC

** End-to-end macrobenchmarks **
NOTE: conv_draw runs on a single core on the benchmark without cross replica sum. A replicated version of conv_draw would not create the fusion nodes found here due to the cross replica sum. As a result, I do not expect a performance hit in typical production use cases.

Speedups (before / after):
INFEED:   0.83x (before  1033.5 ms, after  1250.9 ms): openai_v0_rnn_optimized_df
WINDOW SELECTION:   0.89x (before  8517.4 ms, after  9537.7 ms): conv_draw_batch_128_df
WINDOW SELECTION:   0.91x (before 15847.4 ms, after 17435.4 ms): conv_draw_batch_128_jf
MORE REMAT:   0.93x (before  5367.3 ms, after  5787.3 ms): deeplab_training_img513_depth101_batch_8_jf
WINDOW SELECTION:   0.93x (before  5078.1 ms, after  5453.6 ms): conv_draw_batch_16_df
WINDOW SELCTION:   0.98x (before  7886.3 ms, after  8052.7 ms): conv_draw_batch_16_jf
   0.98x (before 36499.0 us, after 37218.0 us): magenta_jf
   0.99x (before 10003.2 ms, after 10124.5 ms): faster_rcnn_image640_batch_8_df
   0.99x (before  7618.2 ms, after  7687.7 ms): alphachess_rec_df
   0.99x (before 74307.0 us, after 74846.0 us): translate_inference_jf
   1.01x (before 29668.0 us, after 29515.0 us): translate_inference_hybrid_df
   1.01x (before   795.7 ms, after   791.5 ms): deleuze_learn_jf
   1.01x (before   438.1 ms, after   435.6 ms): wavernn_train_jf
   1.01x (before   137.9 ms, after   137.0 ms): resnet_v2_50_batch_16_df
   1.01x (before 13349.9 ms, after 13263.0 ms): nmt_32k_large_df
   1.01x (before 20215.9 ms, after 20083.0 ms): nmt_32k_large_jf
   1.01x (before  2602.9 ms, after  2585.3 ms): openai_v0_rnn_natural_df
   1.01x (before   539.1 ms, after   535.2 ms): deleuze_learn_df
   1.01x (before  3831.9 ms, after  3801.5 ms): nmt_transformer_big_adafactor_with_model_split_jf
   1.01x (before  6510.3 ms, after  6456.9 ms): nmt_model_parallel_jf
   1.01x (before  1538.1 ms, after  1525.2 ms): babelfish_transformer_big_adafactor_with_model_split_jf
   1.01x (before  2384.5 ms, after  2362.3 ms): nmt_transformer_big_adafactor_with_model_split_df
   1.01x (before  6567.9 ms, after  6503.3 ms): nmt_single_tpu_jf
   1.01x (before   285.7 ms, after   282.8 ms): wavenet_train_ar_jf
   1.01x (before 54764.0 us, after 54203.0 us): resnet_imagenet_jf
   1.01x (before   193.0 ms, after   190.9 ms): wavenet_train_ar_df
   1.01x (before 46145.0 us, after 45632.0 us): translate_inference_hybrid_jf
   1.01x (before  4516.0 ms, after  4462.8 ms): nmt_model_parallel_df
   1.01x (before  2071.6 ms, after  2045.9 ms): babelfish_jf
   1.01x (before 61792.0 us, after 60928.0 us): inception_v3_batch_8_train_jf
   1.01x (before  4566.1 ms, after  4499.8 ms): nmt_single_tpu_df
   1.02x (before  6090.6 ms, after  5999.2 ms): alphastar_df
   1.02x (before   233.6 ms, after   230.1 ms): nmt_model_char_v1_df
   1.02x (before 39605.0 us, after 38971.0 us): resnet_imagenet_df
   1.02x (before 29482.0 us, after 28982.0 us): inception_v2_batch_8_train_jf
   1.02x (before   571.6 ms, after   561.8 ms): nmt_model_char_v0_df
   1.02x (before  1173.0 us, after  1153.0 us): smartcompose_single_step_jf
   1.02x (before 11952.1 ms, after 11735.3 ms): faster_rcnn_image640_batch_8_jf
   1.02x (before 22885.0 us, after 22449.0 us): inception_v2_batch_8_train_df
   1.02x (before 31350.0 us, after 30720.0 us): magenta_df
   1.02x (before 10627.0 ms, after 10397.1 ms): alphastar_jf
   1.02x (before 44625.0 us, after 43601.0 us): inception_v3_batch_8_train_df
   1.02x (before 44470.0 us, after 43410.0 us): translate_inference_df
   1.02x (before   699.0 us, after   682.0 us): smartcompose_single_step_df
   1.03x (before   102.0 ms, after 99217.0 us): magenta_dynamic_jf
   1.03x (before 93749.0 us, after 91198.0 us): magenta_dynamic_df
   1.85x (before  5873.3 ms, after  3177.0 ms): object_detection_ssd_mobilenet_v1_300x300_batch_128_df

Geomean before:    680.7 ms
Geomean after:     677.9 ms
Geomean speedup:    1.00x (geomean before / geomean after)

** HBM usage results **

Size reduction (before / after):
   0.92x    (before 5.92 GiB, after 6.42 GiB): unet_training_jf
   0.94x    (before 1.90 GiB, after 2.02 GiB): babelfish_char2feats02_jf
   0.94x    (before 1.99 GiB, after 2.11 GiB): babelfish_char2feats01_jf
   0.97x    (before 1.45 GiB, after 1.50 GiB): nmt_model_parallel_df
   0.97x    (before 1.45 GiB, after 1.50 GiB): nmt_model_parallel_jf
   0.98x    (before 1.59 GiB, after 1.62 GiB): conv_draw_batch_128_jf
   0.98x    (before 1.62 GiB, after 1.65 GiB): conv_draw_batch_128_df
   0.99x    (before 3.29 GiB, after 3.34 GiB): nmt_32k_large_df
   0.99x    (before 3.29 GiB, after 3.34 GiB): nmt_32k_large_jf
   0.99x    (before 3.34 GiB, after 3.39 GiB): deleuze_learn_df
   0.99x    (before 3.35 GiB, after 3.39 GiB): deleuze_learn_jf
   0.99x    (before 1.18 GiB, after 1.19 GiB): resnet_imagenet_df
   0.99x    (before 1.18 GiB, after 1.19 GiB): resnet_imagenet_jf
   1.01x    (before 8.02 GiB, after 7.98 GiB): retinanet_training_img1024_depth50_batch_8_bf16_df
   1.01x    (before 1.15 GiB, after 1.14 GiB): inception_v3_batch_8_train_jf
   1.01x    (before 1.15 GiB, after 1.14 GiB): inception_v3_batch_8_train_df
   1.01x    (before 8.63 GiB, after 8.55 GiB): retinanet_training_img768_depth50_batch_8_df
   1.01x    (before 6.09 GiB, after 6.02 GiB): retinanet_training_img640_batch_8_jf
   1.01x    (before 77.1 MiB, after 76.2 MiB): magenta_jf
   1.01x    (before 6.63 GiB, after 6.55 GiB): retinanet_training_img768_depth50_batch_8_jf
   1.01x    (before 6.09 GiB, after 6.01 GiB): retinanet_training_img640_batch_8_df
   1.01x    (before 76.8 MiB, after 75.7 MiB): magenta_df
   1.02x    (before 8.62 GiB, after 8.45 GiB): inception_v3_batch_128_train_df
   1.02x    (before 241 MiB, after 237 MiB): alphachess_rec_df
   1.02x    (before 6.47 GiB, after 6.34 GiB): inception_v3_batch_128_train_jf
   1.03x    (before 1.08 GiB, after 1.05 GiB): nmt_model_char_v0_jf
   1.03x    (before 1.08 GiB, after 1.05 GiB): nmt_model_char_v0_df
   1.03x    (before 5.19 GiB, after 5.04 GiB): alphastar_jf
   1.04x    (before 5.47 GiB, after 5.25 GiB): alphastar_df
   1.04x    (before 698 MiB, after 668 MiB): conv_draw_batch_16_df
   1.05x    (before 696 MiB, after 665 MiB): conv_draw_batch_16_jf
   1.05x    (before 6.39 GiB, after 6.07 GiB): deeplab_training_img513_depth101_batch_8_jf
   1.07x    (before 243 MiB, after 227 MiB): alphachess_rec_jf

Geomean before:  1.03 GiB
Geomean after:   1.03 GiB
Geomean size reduction:    1.00x (geomean before / geomean after)

*** Reason for rollback ***

Doesn't break speech models anymore

*** Original change description ***

Automated rollback of changelist 204006812

PiperOrigin-RevId: 219905428