[WIP] Add cross_entropy to torchcompile_cat executor

[riccardofelluga]

What does this PR do?

In an attempt to fix #1654 and partially #1552, this PR adds the necessary ops in the torchcompile_cat list to capture HF CausalLMLoss.

Before merging this PR needs testing with other models, I will write a message with the results of the benchmarks.

[IvanYashchuk]

This executor is meant for fusions with a cat operation. Cross entropy is usually quite far from RoPE (which uses cat). How does adding cross_entopy to torchcompile_cat help?

[riccardofelluga]

@IvanYashchuk Adding cross entropy does indeed not help with rope but it allows thunder to use a very efficient fused cross entropy triton kernel which perf currently cannot be matched with the other executors(not even apex). From lines 211 and 212, it does seem that while the set of ops was originally though for rope, it is suggested that other ops could have been added in the future making this a fusion executor that comes in rescue of nvFuser when it can improve performance. Do you think that it would be better to create another executor entry for inductor cross entropy only?