new collective development for JAX/XLA

[i-chaochen]

the initial work: https://github.com/ROCm/mori/tree/pemeliya/collectives-develop

Summary

Add a header-only, CCO-backed collective implementation intended for
XLA integration. The new path uses CCO symmetric windows and SDMA
copy-plus-completion packets for intra-node GPU communication without
depending on MORI's SHMEM collective path.

This PR provides the MORI-side facade and kernels; it does not yet add
JAX/XLA registration or Bazel integration.

Changes

• Add CollectivesFacade for symmetric allocation, lifecycle management,
  kernel dispatch, and runtime tuning.
• Add SDMA push implementations for:
  • reduce-scatter
  • all-reduce
  • all-gather
  • all-to-all
  • collective-permute
• Add a direct-XGMI pull implementation for reduce-scatter.
• Add sliced/pipelined reduce-scatter and all-reduce execution.
• Add SDMA ADD32 completion packets and supporting runtime plumbing.
• Add a unified multi-GPU benchmark with device-side correctness checks
  and runtime selection of collective, size, algorithm, and slice count.

Current scope

• Intra-node LSA/XGMI communication only.
• Real kernels are compiled for gfx942, gfx950, and gfx1250.
• Reduction dispatch defaults to f32 with sum.
• Collective-permute currently supports one destination per rank.
• Back-to-back launches require stream completion and a host-side
  ccoBarrierAll.
• Send, receive, device barrier, quiet, and fence APIs are placeholders.

Validation

The benchmark validates each rank's output after timing.

Before submission, record:

• GPU and ROCm versions:
• PE counts and message sizes:
• Collectives and modes tested:
• Build command:
• Benchmark commands and results:

Performance

Any RCCL comparison should be rerun on the same hardware and software
configuration using the same bandwidth definition. No performance claim
is made from the preliminary checked-in timing log.