aten.native_layer_norm.default lowering modification to include custom compute kernel config

[amalbasaTT]

Ticket

Link to Github Issue

Problem description

ttnn.layer_norm found in albert_v2 models have low accuracy in some occurrences. Adding compute kernel config with fp32_dest_acc_en enabled fixes that issue.

What's changed

1. Added TtnnComputeKernelConfig helper class.
2. Added compute_kernel_config registration to torch.fx graph module.
3. In to_tt_pass.py modified lowering of torch.ops.aten.native_layer_norm.default` to include custom compute kernel config.
4. In tools/generate_op_accuracy_tests.py, expaned _build_code_from_aten_ttnn_graphs to include kernel config at the beginning of forward function.

[kevinwuTT]

I think one cleaner approach when dealing with more complicated calls to ttnn functions is to create a wrapper like this: https://github.com/tenstorrent/pytorch2.0_ttnn/blob/main/torch_ttnn/passes/lowering/target_wrappers.py. That way, we don't have to modify tools/generate_op_accuracy_tests.py because it should automatically grab the wrapper function.

[amalbasaTT]

@kevinwuTT I assumed that in tools/generate_op_accuracy_tests.py we should add variables which are common for most of ops in forward function of generated code (like device for example). compute_kernel_config corresponds to that, and I'm not sure how will wrapping layer_norm make it cleaner (if i add compute_kernel_config to layer_norm wrapper, there will be as many same compute_kernel_configs as there is layer_norms inside graph).