[GlobalOptimization] Rewrite QDQ contractions into integer math (#24902)

Rewrites contractions over `iree_linalg_ext.dequantize_affine` operands
into integer contractions, integer zero-point corrections, and
floating-point scaling.

The transformation uses indexing maps and iterator types to handle named
and generic contractions, including convolutions. It reduces in integer
arithmetic over dimensions where quantization parameters are invariant.
For blockwise quantization, it computes integer partial results, applies
each block's correction and scales, then accumulates across blocks in
floating point.

Zero-point correction terms are built as separate computations so
constant terms can be evaluated independently. Corrections are applied
before conversion to floating point. Scaling and accumulation use at
least f32, preserving wider scale or output types and narrowing only the
final result.

The rewrite requires a plain multiply-add body, zero initialization, and
positive static integer reduction extents whose products and corrections
fit in signed i32. Unsupported cases retain the original computation.
The pass is enabled by default; reassociation changes floating-point
rounding. It can be disabled with
`--iree-global-opt-enable-qdq-to-integer-math=false`.

Builds on the [affine QDQ operations PR
#24901](https://github.com/iree-org/iree/pull/24901) and implements the
contraction rewrite portion of [Linalg(_ext) native QDQ
#24862](https://github.com/iree-org/iree/issues/24862).

---

<sub>Stack created with <a
href="https://github.com/github/gh-stack">GitHub Stacks CLI</a> • <a
href="https://gh.io/stacks-feedback">Give Feedback 💬</a></sub>

---------

Signed-off-by: Thomas Ziereis <ziereis@roofline.ai>
Signed-off-by: Florian Walbroel <walbroel@roofline.ai>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Florian Walbroel <walbroel@roofline.ai>