torch.aten.batch_norm is relying on incorrect annotations

Nobody has claimed this yet.

As discussed in https://github.com/pytorch/pytorch/issues/73050#issuecomment-1051382044, there are a few ops that don't correctly annotate that they mutate their operands. It seems like those are `aten::batch_norm` and `aten::layer_norm`.

When I revamped our ODS generator code, I tried correcting those exceptions w.r.t. the HasValueSemantics and ReadOnly traits, but it seems we were relying on the old, incorrect annotation (which I think was okay, since it only matters in the training case, which we haven't implemented yet)
https://github.com/llvm/torch-mlir/blob/a5fe0cf06308af3a372e40ae927995f7920fb55d/python/torch_mlir/dialects/torch/importer/jit_ir/build_tools/registry.py#L258

To work on this, you just have to uncomment the code linked above and regenerate the ODS for `torch.aten.batch_norm` and see what breaks in the tests. I dug into it a little bit, and it seems like we will need some special handling in ReduceOpVariants to convert torch.aten.batch_norm to value semantics when `training == false`.

Start with the linked registry.py code and uncomment the annotation changes for torch.aten.batch_norm and torch.aten.layer_norm. Regenerate the ODS and run the affected tests, then inspect ReduceOpVariants for the required special handling when training == false. Done means the regenerated annotations and tests work without breaking existing behavior.

Written by the indexing model from the issue text.