Add fp16 support for BatchNormalization Forward/Backward

[mindest]

Description

• Add fp16 support for BatchNormalization ForwardTraining/Backward
• Bind inputs mean and var with outputs mean and var, respectively

Motivation and Context

• Currently BatchNormalization does not support fp16 and should be kept in fp32 when training
• Fix the problem that running_mean and running_var do not change in training mode

[pranav-prakash]

Note that as I mentioned in #6946 the cuda kernel for BatchNorm breaks the ONNX spec in outputtting inverse standard deviation instead of variance for the saved statistic. While doing so avoids the need for an extra operation here since we can use the cudnn output directly, I'm not sure if it's worth breaking spec for that (and that too it's probably not a significant performance gain).

But then again the spec for batchnorm was recently updated to drop the redundant "saved mean/var" output (onnx/onnx#3379), so I think those two outputs can just be removed entirely (at least for opset 14).

[SherlockNoMad]

Hi @mindest, which model need this op? is this urgently needed?