Batchnorm training mode support in a minimal build

[baijumeswani]

Now that ONNX Runtime supports a minimal training build, the batchnorm in training mode is needed even in a minimal build.

This PR removes the preprocessor macro BATCHNORM_INCLUDE_TRAINING_SUPPORT and replaces it with ENABLE_TRAINING_OPS

[baijumeswani]

Thanks for the review @edgchen1 :)

[BowenBao]

This is excluding batchnorm training mode support for standard onnxruntime releases, regression found from ort-nightly==1.16.0.dev20230821001. Can we revert that effect? There are some exported models for inference that runs just bn specifically in training mode.

[baijumeswani]

@edgchen1 @skottmckay
Is it ok to completely remove the need for the preprocessor guard for training mode batchnorm? This might impact the inference minimal build.

@BowenBao under what conditions would you need the training batchnorm for inferencing?

[baijumeswani]

I created #17270

But I am not sure I understand why a training op would be needed in an inference build. @BowenBao could you use onnxruntime_ENABLE_TRAINING_OPS cmake flag in your test if needed or do we need the training bits in the inference build?

[BowenBao]

afaik some models are set to manipulate running mean and running variance in some ways. There is a separate issue pytorch/pytorch#75252 discussing other ways of export but there is no traction.

It would be nice if the previous behavior can be restored for 1.16 release. I hope training minimal build would still have this op after #17270 so the goal of this PR is not affected?