fix: derive ep_rank from inner_rank / mlp_tp_size instead of inner_rank % ep_size

[Copilot]

Motivation

When both mlp_tp_size > 1 and ep_size > 1, computing ep_rank = inner_rank % ep_size collides with mlp_tp_rank = inner_rank % mlp_tp_size — both consume the same modulo space, producing duplicated/incorrect EP ranks. The invariant inner_rank == ep_rank * mlp_tp_size + mlp_tp_rank must hold.

Modification

• src/turbomind/turbomind.cc: Fix ep_rank derivation to inner_rank / mlp_tp_size; fix mlp_tp_rank to inner_rank % mlp_tp_size; restore assertion to mlp_tp_size * ep_size == comm_size_; add GetEpRank() implementation
• src/turbomind/turbomind.h: Declare GetEpRank(int index)
• src/turbomind/models/llama/llama_params.h: Add int ep_rank = 0 to EngineParam
• src/turbomind/engine/engine_config.h: Add ep_size (default 1) to EngineConfig
• src/turbomind/python/bind.cpp: Expose ep_rank via Python binding

The correct decomposition:

inner_rank = ep_rank * mlp_tp_size + mlp_tp_rank
           ⟹  mlp_tp_rank = inner_rank % mlp_tp_size
               ep_rank     = inner_rank / mlp_tp_size

BC-breaking (Optional)

No BC-breaking changes. ep_size defaults to 1, preserving existing single-EP behavior.

Use cases (Optional)

N/A

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
4. The documentation has been modified accordingly, like docstring or example tutorials.