[GlobalOpt] Keep strided input maps when the producer can fuse - #24950

Factoring `dim * cst` strides out of a contraction-like generic inserts a
strided `tensor.extract_slice` between the contraction and its input. When
the input has an elementwise producer (a dequantize feeding a 1x1 stride-2
convolution, say), the slice blocks both fusion paths: elementwise fusion
requires the producer to directly define the consumer's operand, and
tile-based producer fusion only supports unit-stride slices. Codegen falls
back to materializing the producer as a whole-tensor stack allocation in
every workgroup: a 3.2 MB alloca for a 1x256x56x56xf32 dequantize, well
over the 32 KB stack limit, and the compile fails.

Skip the factoring only when the strided input's producer can fuse
elementwise into the contraction and into every other user of its result,
so the fusion actually eliminates the producer. Codegen then merges the
producer into the contraction, which reads the input directly through the
strided map with no intermediate tensor. Producers that are materialized
anyway (e.g. a result also returned from the function) keep the factored
projected-permutation input map.

Strided contractions with a fusable producer now lower through the generic
path instead of contraction selection such as AMDGPU MFMA, the original
motivation for the factoring, in exchange for eliminating the intermediate
tensor. Strided contractions on dispatch-boundary inputs keep the factored
form.

Fixes iree-org#24752

Tests:
- iree-opt --split-input-file --mlir-print-local-scope
  -iree-global-opt-convert-strided-contraction-to-contraction
  compiler/src/iree/compiler/GlobalOptimization/test/strided_contraction_to_contraction.mlir

Co-authored-by: OpenAI <noreply@openai.com>
Signed-off-by: zhangpan <98025960+zhangpan2001@users.noreply.github.com>