[Refactor][TCPCG] Migrate MoE and Mamba graph execution to eager methods (6/9)

[Oasis-Git]

This is step 6/9, split from the original TCPCG removal PR #41634.

Depends on #42289 (5/9); merge those dependencies first. This branch targets upstream main, so its diff includes unmerged prerequisites. Review this step alone using the isolated step-6 diff.

Series

1. Relocate shared graph tensor and DSA head-gate helpers — [Refactor][TCPCG] Relocate shared graph tensor and DSA head-gate helpers (1/9) #42285.
2. Add explicit batch support to BCG eager regions — [Refactor][TCPCG] Bind live forward batches in BCG eager calls (2/9) #42286.
3. Retire TCPCG backend selection, configuration, documentation, and tests — [Refactor][TCPCG] Retire backend selection and configuration (3/9) #42287.
4. Consolidate Radix attention eager regions and isolate Inkling-specific behavior — [Refactor][TCPCG] Consolidate attention eager regions and isolate Inkling policy (4/9) #42288.
5. Migrate DSA and DeepSeek eager regions — [Refactor][TCPCG] Migrate DSA and DeepSeek graph breaks to eager methods (5/9) #42289.
6. Migrate MoE and Mamba eager regions — this PR.
7. Migrate diffusion eager wrappers. Depends only on step 2.
8. Remove obsolete runtime contexts, registries, and split-op plumbing.
9. Simplify quantization paths separately because of their torch.compile implications.

The disputed relocation of inactive compiler/reference files is excluded pending a separate decision.

Changes

Move the remaining MoE and Mamba graph execution away from TCPCG split-op adapters and hidden batch/module lookups.

• Use a class-owned eager method for expert all-to-all, preserving its capture stub and stable output buffer.
• Run Nemotron Mamba through an eager method with an explicit batch, real-token slicing, padded-tail zeroing, and replay-time collective policy restoration.
• Remove TCPCG-only fused-MoE dispatch and the Qwen3-Next TC-specific normalization choice.

Runtime registries remain until the cleanup step. This step uses the batch binding introduced in step 2.

Validation

Pre-commit and Python syntax checks passed on the changed files. Distributed MoE collectives, Mamba CUDA replay, model accuracy, and full main CI at this intermediate step remain to be validated; local macOS checks do not establish accelerator correctness.

---

CI States

Latest PR Test (Base): ❌ Missing run-ci label -- add it to run CI tests.
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.