Need MLIR-level MC2 collective / comm-context surface and better 8-rank routing diagnostics

[zhoubot]

Summary

PTOAS can already lower local routing primitives like TGATHER / TSCATTER, but MC2 routing ports still have no MLIR-level surface for the distributed communication context itself.

Current state

In the local 910B migration workspace:

• moe_distribute_dispatch and moe_distribute_combine now compile through PTO-DSL -> PTOAS -> bisheng for the real A2 contract (epWorldSize=8, H=7168).
• PTO source is explicit-sync-free and PTOAS insert-sync is active.
• The actual 8-rank routing benchmark still times out before any rank report is emitted, so failures are currently only visible at the outer harness level.

Gaps exposed by bring-up

• No MLIR-level op or abstraction for HCCL / parallel-group / window context used by MC2 collectives.
• No direct way to represent the communication boundary in PTO IR, so MC2 ports fall back to host-managed HCCL steps outside PTOAS.
• Diagnostics for multi-rank routing failures are too indirect; once ranks stall, PTOAS provides no comm-specific breadcrumbs to distinguish context/setup issues from lowered-kernel issues.

Requested work

• Add PTO IR / lowering surface for MC2 collective context and collective-style operations.
• Improve diagnostics around comm-lowered kernels so 8-rank routing failures can be attributed earlier and more precisely.
• Keep A2/A3 autosync behavior explicit for scalar-pipe comm instructions such as TSCATTER.