Add Fp16 fused lstm w/ Eigen (on CPU) and add fp16 unittests

* Add unittests for fp16 fused lstm and cell_clipping
  * Init weights with 6 non-zero mantissa bits
  * Pick rtol, atol = 8e-3, 8e-4 (precision for values in [0, 0.5] is about 5e-4)
* Add fp16 benchmarks and benchmark target

PiperOrigin-RevId: 219877548