Pass the device ordinal to use for execution to the XLA compiler for

auto-tuning.

Previously, when compiling a graph for multiple devices concurrently, XLA
would use the default device for auto-tuning. With this patch tf_cnn_benchmark
with model resnet50 finishes on 8 V100s batch 128, and gets a speedup of ~20%
over a single one; the next steps are to get it to run at batch 256 and to
scale well.

PiperOrigin-RevId: 206720140