[Torch] Preserve PT2E quantization as LinalgExt ops - #24918

Preserves PT2E quantization during Torch input conversion as
iree_linalg_ext.quantize_affine and
iree_linalg_ext.dequantize_affine operations. This retains the affine
parameters needed by the propagation and integer-contraction rewrites instead
of expanding quantization into scalar arithmetic during Torch-to-Linalg
lowering.

The conversion handles the typed per-tensor and per-channel PT2E quantize and
dequantize operations. It carries constant quantization bounds, storage
signedness, optional zero points, and scalar or channel parameter maps into the
LinalgExt representation. Operations with dynamic bounds, dynamic axes, or
incompatible tensor metadata remain available to the general torch-mlir
lowering.

Builds on the QDQ propagation in #24906
and requires the typed PT2E operations supplied by the torch-mlir update in
#24908. Until #24908 lands, its
one-line submodule update also appears in this stacked diff. This completes the
Torch import portion of Linalg(_ext) native QDQ #24862.

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}