[None][feat] WideEP FT: add EPLB mask-only reconfigure

[chienchunhung]

Summary

• Add a C++ MoeLoadBalancer::reconfigureMaskOnly(deadRanks) path that rebuilds EPLB placement metadata while leaving weights in place.
• Keep dead ranks excluded from later dynamic EPLB placement/replication so degraded routing does not drift back onto failed ranks.
• Expose the API through nanobind and the Python MoeLoadBalancer wrapper, with focused C++/Python tests.

Why

1b.1 needs EPLB to make dead-rank slots unreachable after the communication-layer active-rank mask is installed. This implements the MVP mask-only slot remap path and fails closed if any expert would lose its last surviving replica.

Validation

• git diff --check
• python -m py_compile tensorrt_llm/_torch/modules/fused_moe/moe_load_balancer.py tests/unittest/_torch/modules/test_moe_load_balancer.py
• Attempted: PYTHONPATH=. pytest -q tests/unittest/_torch/modules/test_moe_load_balancer.py -k 'lifecycle_methods or reconfigure_mask_only_rejects_active_iteration' (blocked locally: torch is not installed in this environment)

[chienchunhung]

Superseded by NVIDIA#15525.