[VMVX] Enable scf.forall distribution for VMVX pipelines. (#23615)

The iree_codegen.query_tile_sizes op was resolved during scf.for
distribution. If we go with scf.forall distribution, it becomes
`iree_codegen.ukernel.generic` before reconcile+resolve runs. Thus, it
requires a workaround similar to the existing one.

Further more, LowerUKernelOpsToCallsPass would lower it to a func.call
which is memory-effecting and would be rejected by backward slice
analysis. So we put the reconcile+resolve in between, which is the stage
that we have enough information.

- Move reconcile+resolve before LowerUKernelOpsToCallsPass in the VMVX
variant pipeline.
- Extend the WAR hack in materializeSliceFromOrdinals (TODO #13038) to
also handle the lowered UKernelGenericOp form of query_tile_sizes,
replacing it with a constant 16 in the workgroup count region.
- Disable the distribution if it identifies that the distribution is
already applied.

Fixes https://github.com/iree-org/iree/issues/23588

---------

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