[Codegen][PCF] Add foldForallIntoPCFLoop for split-k loop handling (#23452)

Adds a generic `foldForallIntoPCFLoop` function that folds an
`scf.forall` containing a `pcf.loop` into a single `pcf.generic`
operation. This enables incorporating user expressed extra levels of
parallelism into a scope. The immediate use case this enables is split-k
by folding the split-k loop into the workgroup loop.

Key changes:
- Add `foldForallIntoPCFLoop` API in PCF/Transforms/Transforms.h
- Implement structural matching helpers (matchFoldTerminator,
matchFoldPCFLoop, matchFoldWriteSlices) that validate requirements
without scope-specific logic
- Add TestFoldForallIntoPCFLoopPass for unit testing with local_mapping
+ sequential scope
- Add FoldSplitKWorkgroupLoop pattern in ConvertWorkgroupForallToPCF.cpp
that matches split_reduction_mapping + workgroup_scope and calls the
generic fold
- Run fold pattern as second pass in ConvertWorkgroupForallToPCFPass

The fold operation:
1. Creates pcf.generic with same scope as inner pcf.loop
2. Linearizes forall iteration space and delinearizes inside generic
3. Converts pcf.loop to a nested scf.forall loop to handle spillover
4. Composes tensor.parallel_insert_slice with pcf.write_slice ops

The reason this pattern has to generate a pcf.generic instead of
pcf.loop is to avoid extra execution of code that was inside
the scf.forall but not inside the pcf.loop.

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>