Add verifier for AtenMatmul to check that 1D x 1D outputs a scaler

[weiyu0824]

I'm adding a rank verifier for the AtenMatmul operation. The goal is to catch rank mismatches earlier in the pipeline.

Problem

Currently, invalid result rank error only show up during the lowering stage (e.g., when running --convert-torch-to-linalg) and the error (torch.cast error shown below) doesn't clearly point to the root cause.

Solution

Adding a verifier to report a clear, specific error directly on the AtenMatmul op itself, before lowering even starts.

• Case: 1D x 1D
  I add a check for 1D × 1D Matmul, as Pytorch Matmul doc define that 1D x 1D must output a scalar (rank 0). The verifier ensures the output rank is correct.

Before and After

If we feed an invalid (1D x 1D = 1D) MLIR like this:

func.func @invalid_matmul_1d_1d(%arg0: !torch.vtensor<[4],f32>, %arg1: !torch.vtensor<[4],f32>) -> !torch.vtensor<[1],f32> {
  %0 = torch.aten.matmul %arg0, %arg1 : !torch.vtensor<[4],f32>, !torch.vtensor<[4],f32> -> !torch.vtensor<[1],f32>
  return %0 : !torch.vtensor<[1],f32>
}

• Before: running torch-mlir-opt input.mlir --convert-torch-to-linalg, we'll get "Error: 'tensor.cast' op operand type 'tensor' and result type 'tensor<1xf32>' are cast incompatible" during lowering.
• After: Running torch-mlir-opt input.mlir, we'll get "Verifier-reported error: "1D x 1D matmul should produce a scalar (rank 0)" during IR verfication

Note: I also added the above-metioned mlir to invalid.mlir, and ninja check-torch-mlir would complete without error.