#9709: Add optional transpose_a and transpose_b to ttnn matmul and linear

[TT-BrianLiu]

Ticket

Link to Github Issue

Problem description

In GGML, inputs a and b to matmul may be "pre-transposed", so to perform matmul properly we need a corresponding transpose a or b.

What's changed

Easiest solution is to add an optional transpose op before running matmul. I added transpose_a and transpose_b flags to specify whether to run the transpose on the inputs or not. This should work for most interleaved cases. As a side note, I also had to fix transpose to respect padding when swapping dims during compute_output_shapes.

There are two concerns that can be addressed down the line:

• Perf: We can technically support the optional transpose directly in the matmuls, but this is a lot more work and requires more testing. But it will give us better perf.
• Functionality: Using composite ops to achieve transpose will limit the input specs available to matmul to what's supported by our transpose op. We can either uplift transpose to support what matmul supports, or if we did the transposes internally in matmul this problem will automatically go away.

Checklist

• Post commit CI passes
  • All post-commit: https://github.com/tenstorrent/tt-metal/actions/runs/9764203272
  • Nightly fast dispatch:https://github.com/tenstorrent/tt-metal/actions/runs/9764205780
  • ttnn unit tests: https://github.com/tenstorrent/tt-metal/actions/runs/9764213369
• Model regression CI testing passes (if applicable)
  • Models post-commit: https://github.com/tenstorrent/tt-metal/actions/runs/9764208694
• New/Existing tests provide coverage for changes