[Codegen][CPU] Run HoistInnerTiledAccReshapes before vector lowering. (#24491)

Mirrors the GPU pipeline (`buildLLVMGPUCodegenPassPipeline` adds this
pass right before `UnrollToIntrinsicsPass`). Before this change, on CPU
the reshape ops left around `iree_codegen.inner_tiled` sat between the
K-reduction loop's `scf.for` iter args and the op's accumulator operand,
so the per-tile `HoistableConversionOp` pairs that the inner_tiled
lowering wraps around its ACC distribute/reassemble couldn't satisfy
`HoistConversionFromLoopPattern`'s strict iter-arg match. They failed to
hoist, got inlined in place, and
`EliminateHoistableConversions` fired a `hoistable_conversion was not
hoisted or cancelled; inlining in place` remark on every CPU matmul
build using the inner_tiled path.

Running `HoistInnerTiledAccReshapes` before
`LLVMCPUVirtualVectorLoweringPass` wraps the prefix/suffix reshape ops
(transpose, shape_cast, broadcast) around the inner_tiled as
`acc_reshape_to_intrinsic` / `acc_reshape_from_intrinsic` hoistable
conversions and runs the elimination over them. The reshapes hoist out
of the K loop, the inner_tiled's accumulator is fed directly by the iter
arg, and the per-tile conversion pairs emitted later in the vector
lowering see direct iter-arg inputs and hoist cleanly.

Progress towards #24323.

Signed-off-by: Benoit Jacob <jacob.benoit.1@gmail.com>