[sharktank] Export Attention IRs for LLMs

[archana-ramalingam]

Export Attention IRs for paged causal and non causal LLMs.

[sogartar]

We can't use directly torch ops with our tensor types.
https://github.com/nod-ai/SHARK-Platform/actions/runs/11148274191/job/30984526213?pr=175#step:6:123
In this example torch.concat would need to be substituted with sharktank.ops.cat.
At some instances we may be lacking the equivalent op or it may have no implementation for certain tensor types, like ReplicatedTensor. In this case we have to add it. For some tensor types this may be difficult, but in most cases it is easy. I have been meaning to make a general implementation when all tensors are replicated across devices, but have not gotten to it. I will add the cat overload for the replicated tensor.