Add conversion for StableHLO scatter op to TTIR and TTNN dialect

[ajakovljevicTT]

As part of the effort to run stableHLO models on TT silicon, lowering scatter op through TTIR and TTNN dialects.
This op currently has limited functionality, but enough to pass the current models, primarily due to the current limitations in tt-metal:

• The op only supports one index, mimicking Torch's select_scatter op https://pytorch.org/docs/stable/generated/torch.select_scatter.html

• Furthermore, since we do not currently support any indexing on the tt-metal side (Limitations in TT-Scatter Op tt-metal#4294), it will ignore the ttir index variable and use index 0, as supported by tt-metal.

• The StableHLO scatter op also supports adding a custom function for merging the two tensor operands of the scatter. Since this is not supported in tt-metal, I have added the check to see if the function is just mapping from one tensor to another (the default select_scatter behaviour) and assert on that. I have opened a related tt-mlir issue (Support of custom functions for the scatter op in TTNN dialect. #1278), but not a tt-metal one.

• Relevant TTIR dialect issue: Add ttir.scatter op to the TTIR dialect #1325

• Edit:
  As advised in the comments changed the runtime pipeline of the scatter op to go through the composite_binary path, omitting any additional (currently not used because they are not supported by tt-metal) op parameters and attributes.

[ajakovljevicTT]

Hi all, would appreciate if someone has the time to take a look at this PR, as it is blocking on some of the models that we are running on tt-torch. Thanks!