[DispatchCreation] Tighten scatter-skip predicate in CollapseDimensions (#24334)

PR #24034 skipped collapse for any `linalg.generic` with a
`tensor.extract` and a `linalg.index` in its body, in order to avoid an
expensive delinearization on strided-scatter generics produced by
`ConvertStridedInsertSliceToGeneric`. That predicate was too broad and
caught the RoPE + FP8 dispatch in Llama-8B, whose fused body also has
`tensor.extract` + `linalg.index` (rotate_half lookup) but is not a
strided scatter; the collateral block prevents the `4×batch×seq` outer
collapse and produces a 5-D dispatch (`4x2048x8x2x64`) that tiles ~3.5x
slower on gfx942 than the pre-#24034 `8192x8x128` shape.

Tighten the predicate to require all three signals of the
strided-scatter shape: empty `ins` operands (source captured via
`tensor.extract`), at least one `linalg.index`, and a `tensor.extract`
whose result is consumed by an `arith.select` (the bounds check).

Fixes: https://github.com/iree-org/iree/issues/24322

---------

Signed-off-by: yzhang93 <zhyuhang88@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>