Refactor EP runtime architecture

[Binyang Li (Binyang2014)]

Unify latency and overlap under one runtime, organize kernels by dispatch and combine algorithms, and expose one opaque dispatch handle.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Binyang Li (Binyang2014)]

Addressed the latest review in 2090bd5:

• Backend now uses one standard super().__init__(config, output_layout) path; there are no direct class __init__ calls and no separate mode instances.
• Removed get_combine_input_buffer(). Rank-major dispatch returns the runtime-owned MLP/combine destination as dispatch_out.combine_input_buffer, tying it to the dispatch result and active layout.

[Binyang Li (Binyang2014)]

Updated backend hierarchy in db0cb30: Backend is now an abstract parent, create_backend() selects LatencyBackend or OverlapBackend from MoEMode, and each concrete backend owns the same unified C++ Runtime type. The previous multiple-inheritance/MRO initialization is removed.

[Binyang Li (Binyang2014)]

Updated in 5d752e3: all internal Python imports under mscclpp.ep now use full absolute package paths.

[Binyang Li (Binyang2014)]

Updated in 61eb952: create_backend() and concrete backend constructors now accept only MoECommunicatorConfig; each backend reads and resolves config.output_layout directly. The duplicate layout plumbing is removed.

[Binyang Li (Binyang2014)]

Updated in e7fb586: runtime-owned generic, BF16, and FP8 CUDA pointer tensor views are now centralized in mscclpp.ep.utils; latency.py only selects layout/dtype and owns mode behavior.

[Binyang Li (Binyang2014)]

Updated in f4377dd: removed the latency-only dispatch output getter. Every mode now exposes dispatch output uniformly through dispatch_out.tokens; runtime allocation remains internal.

[Binyang Li (Binyang2014)]

Updated in 6e44bb7: dispatch datatype resolution and scale metadata helpers are now common mscclpp.ep.utils functions. latency.py retains only latency-mode support policy and execution.

[Binyang Li (Binyang2014)]

Updated in a67f794: BF16 and FP8 pointer wrappers are merged into one dtype-driven tensor_from_pointer() helper, also used for INT32/FP32 metadata.

[Binyang Li (Binyang2014)]

Updated in 5125a93: removed the redundant bf16_view() wrapper; overlap mode now calls the generic dtype-driven tensor_from_pointer() directly.

[Binyang Li (Binyang2014)]

Updated in 4faa42c: removed LOW_LATENCY and HIGH_THROUGHPUT compatibility aliases. MSCCL++ now exposes only MoEMode.LATENCY and MoEMode.OVERLAP.

[Binyang Li (Binyang2014)]

Updated in 3c9ef8b: removed mode suffixes throughout the EP surface. Fixed-buffer algorithms use qualifier-first names (latencyExpertMajor, latencyRankMajor, etc.); token-major APIs use layout-first names (tokenMajorPrepare, tokenMajorDispatch, tokenMajorReduce).

[Binyang Li (Binyang2014)]

Updated in 5b8ae8a: api.cuh now documents each expert-major, rank-major, and token-major dispatch/combine contract separately. DeviceContext also carries its persistent devicePtr_, so every API accepts one context argument instead of separate host/device context parameters.

[Binyang Li (Binyang2014)]

Updated in 21f74b9: flattened kernel sources directly under dispatch/ and combine/; backend/runtime selection now owns mode policy. Fixed-buffer APIs use expertMajorDispatch/rankMajorDispatch, while token-major remains tokenMajorPrepare/tokenMajorDispatch/tokenMajorReduce.

[Binyang Li (Binyang2014)]

Updated in 654e707: split the generic combine API into layout-specific entry points: expertMajorRankLocalReduce, rankMajorGatherReduce, and expertMajorDirectSend.

[Binyang Li (Binyang2014)]

Updated in f0b8759: all combine kernels and entry points now use a Combine suffix, matching dispatch naming (expertMajorLocalReduceCombine, rankMajorGatherReduceCombine, expertMajorDirectSendCombine, tokenMajorReduceCombine).

[Binyang Li (Binyang2014)]

Updated in 5024680: replaced latency_num_blocks and num_sms tuning with one optional num_blocks field (defaults: 130 latency, 20 overlap). Test-only NCCL monitoring overrides were also removed; hardware SM-count fields remain internal.