Remove _log_softmax/_softmax in favor of log_softmax and softmax respectively. #76433
Description
🚀 The feature, motivation and pitch
Similar rationale to #73135
The only difference between _softmax and softmax is that softmax takes an optional dtype argument, while _softmax takes a half_to_float argument. However, this half_to_float arg is not needed in the backwards pass, and even if it were, we could simply reconstruct it from the argument dtypes.
https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/SoftMax.cpp#L427
So, to remove these, we can simply turn the current _softmax and _log_softmax into C++ functions (not operators), and mostly preserve the same logic as today. The backwards pass would need to be modified slightly to take account of the cast.
cc: @albanD @zou3519 @ngimel @ezy
Alternatives
No response
Additional context
No response