Gather op implementation [#1015]

[ddilbazTT]

Gather op implemetation end-to-end.

SHLO GatherOp --> TTIR Gather Op --> TTIR Embedding Op (+ if needed, Reshape Op) --> TTNN Embedding Op (+ if needed, Reshape Op)

TTNN does not currently support Gather Op, but had Embedding Op support.
Given that Torch Embedding is lowered into SHLO Gather Op, and that most
Gather Ops in models are simple embeddings, Gather Op is lowered into
Embedding Op. If more sophisticated Gather Op implementations are
encountered, then Gather could be lowered into slice/ concat/ etc if
needed.

Gather Op checks certain criteria to ensure lowering into Embedding Op
in lib/Conversion/TTIRToTTIRDecomposition/TTIRToTTIRDecomposition.cpp

• Output tensor shape: Multi-dimensional with last dimension as
  embedding size/ hiddenDim
• Slice sizes: Must be [1, hiddenDim], where hiddenDim matches last
  output dimension
• Offset dimensions: Strictly [2]
• Collapsed slice dimensions: Strictly [0]
• Start indices shape: Must be compatible with output shape
  • startIndices.size() < output.size()
  • if startIndices.size() == output.size(), then startIndices[-1] == 1
  • Last dimension of start indices can be reduced by reshape op.
  • • This is due to embedding weights requiring to have smaller
      size than output shape

[rpavlovicTT]

Few additional comments

[ddilbazTT]

Thanks everyone for your feedback! I hope the latest version addresses everybody's concerns.