[CPU] Improve splitReduction for bounded-dynamic cases. (#24067)

The revision reworks the upstream `splitReduction` within IREE and
supports the bounded-dynamic cases.

The structure of splitReduction function mirrors the inner-parallel path
of upstream `mlir::linalg::splitReduction` as closely as possible to
keep the port easy to maintain. The only intentional deltas are:
1. Uses `OpFoldResult` + `tensor::getMixedSize(s)` so non-reduction dims
and the outer reduction dim can be dynamic.
2. Computes the outer reduction size with `affine.apply floorDiv
splitSize` instead of dividing a static shape at compile time.
3. Passes `output_shape` to `tensor.expand_shape` so the outer dim can
be a dynamic value.
4. Single-input / single-output only (no outer loop over input
operands).
  5. Inner-parallel only (no outer-parallel branch, no `useAlloc`).

Fixes https://github.com/iree-org/iree/issues/24010

---------

Signed-off-by: hanhanW <hanhan0912@gmail.com>