Unnecessary buffers creation when enabling scf.forall distribution on the vector distribute pipeline.

[pashu123]

Dump with forall distribution: https://gist.github.com/pashu123/2a162391c5212dc7351a08d0748833fd

Dump without forall distribution: https://gist.github.com/pashu123/b49299b19d14959244079d75ecc502ba

After the IR dump after EmptyTensorToAllocTensor pass the dump with forall distribution has an extra buffer.

%17 = bufferization.alloc_tensor() : tensor<64x64xf16>
%31 = vector.transfer_write %30, %17[%c0, %c0] {in_bounds = [true, true]} : vector<64x64xf16>, tensor<64x64xf16>
%extracted_slice = tensor.extract_slice %arg3[%arg0, %arg2, %arg1, 0] [1, 64, 1, 64] [1, 1, 1, 1] : tensor<2x4096x10x64xf16> to tensor<1x64x1x64xf16>
%inserted_slice = tensor.insert_slice %31 into %extracted_slice[0, 0, 0, 0] [1, 64, 1, 64] [1, 1, 1, 1] : tensor<64x64xf16> into tensor<1x64x1x64xf16>

Looking further, the allocated tensor (%17) seems to be redundant, and we can directly transfer write to the %extracted slice.

[pashu123]

Reference PR: #19420