[PyTorch] Achieve working fp16 support

[mwootton]

Identify and fix outstanding issues with fp16.

[iotamudelta]

Note that the HIP fixes bring us somewhat closer to working fp16. @lcskrishna tagged here as he knows the current status for overall training - please comment.

[lcskrishna]

currently there are two issues that i identified for training with fp16 so far:

1. batchnorm-fp16 doesn't go through miopen path. That needs to be enabled. Also, it generates the NaNs in the outputs - ignoring the path stuff.
2. max/min calls are getting failed, that stops the training after first epoch.

[iotamudelta]

As of ROCm 1.9.2, this now works.