Support merge_fp32_inputs_into_fp16 for predefined partitions (#35361)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35361
If the inputs we are bundling together will be consumed by ops from the same partition, we can assign the Split and Half2Float ops to the that partition too. Otherwise, we do nothing.
Reviewed By: bangshengtang
Differential Revision: D20639777
fbshipit-source-id: 4032abb9178f3b44a85e4789ddf5ad5624245e3a
