Tracking issue: PyTorch precision upcast issue.

[ysiraichi]

This issue tracks the development of the output upcasting issue present in PyTorch/XLA.

• Main issue: Unexpected upcasted output when XLA_USE_FP16 and XLA_USE_BF16 are set. #6403
• Temporary fix for the benchmarking scripts: Force to fp16 models to fp32 if XLA_USE_FP16 is already set. #6389
• PyTorch nn.Module conversion fix: Add swap_tensors path to nn.Module._apply pytorch#117167

Once pytorch#117167 lands, we should revert pytorch/xla#6389, and stop setting XLA_USE_{FP16,BF16} in the benchmarking script.

cc @miladm @JackCaoG @golechwierowicz