[GPU] Allow swizzle promotion for transpose linalg.generic producers (preshuffling prs 1/3) (#24086)

This PR is the first in a series of PRs meant to enable preshuffling for
scaled gemms.

Currently, operands to be swizzled are skipped when producing
SwizzleHintOps if their producer is a linalg::LinalgOp. This means that
transpose linalg.generic ops feeding into a matmul with SwizzleOperand
promotion would never get XOR swizzle hints applied to their output
buffer.

This change adds an exception: transpose linalg.generic producers
(identified via linalg::isaTransposeOpInterface) no longer exit early.
For a GEMM dispatch where the transpose is folded into memory access,
this does not produce any extra allocations, bufferization correctly
recognizes that the `tensor.empty` associated to the TransposeOp and the
SwizzleHintOp are redundant and removes the copy.

Non-transpose linalg.generic producers continue to get the early return
with lowering_config annotation as before.

Made-with: Cursor

---------

Signed-off-by: Muzammiluddin Syed <muzasyed@amd.com>