feat(lingbot-video): add native Dense/MoE inference and four-GPU runtime

[lzx1413]

Description

Add native TeleFuser support for LingBot-Video Dense 1.3B and MoE 30B checkpoints.

The PR provides single-process and four-GPU inference for T2I, T2V, and TI2V,
including the MoE low-noise refiner. It preserves upstream conditioning and
sampling semantics while exposing TeleFuser-native service, parallelism, and
validation workflows.

Motivation

LingBot-Video previously required the upstream runtime and did not have a
TeleFuser pipeline, service contract, or validated distributed execution path.

This change makes LingBot-Video available through the standard TeleFuser
pipeline/service interfaces, supports four-GPU inference without wrapping the
service in torchrun, and keeps upstream-compatible behavior for structured
captions, VAE latents, visual conditioning, and base-to-refiner handoff.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Performance improvement
• Code refactoring
• Documentation update
• Other (please describe):

Changes Made

• Add LingBot-Video Dense and MoE DiT implementations, checkpoint loaders, Qwen3-VL
  structured-caption encoding, Wan VAE stages, scheduler integration, and T2I/T2V/TI2V pipelines.
• Add native MoE refiner execution with in-memory RGB handoff; preserve VAE
  normalization, TI2V visual and clean-latent conditions, geometry handling,
  prompt envelopes, and upstream-compatible default negative conditions.
• Add Dense and MoE service examples, contracts, CLI entry points, default
  TF_MODEL_ZOO_PATH model locations, and user-facing example documentation.
• Add four-GPU FSDP/Ulysses execution, worker-owned scheduler loops, grouped-GEMM
  MoE routing, CFG2×SP2 support, batch-CFG conflict validation, and per-rank CPU
  thread limits.
• Add opt-in FP8 W8A8 expert execution via expert_backend=fp8; grouped BF16 GEMM
  remains the throughput default.
• Keep base and refiner workers co-resident for four-GPU refinement, then release
  them before high-resolution VAE decode; retain the lower-memory serial fallback.
• Add benchmark, capture, replay, parity, distributed/SP, checkpoint inspection,
  and base-to-refiner handoff validation tools with unit coverage.
• Add regression fixes for distributed refiner handoff restoration, grouped-MoE
  FP32 route reduction, CFG behavior at guidance scale <= 1, refiner constraints,
  prompt-duration precedence, and output suffix handling.
• Add source attribution and update README, service, configuration, parallelism,
  and project guidance documentation.

Testing

• Targeted LingBot-Video unit/model/pipeline/tool tests pass
• Manual GPU testing performed
• Benchmarks and validation tooling added

Test commands:

ruff check <changed-python-files>
ruff format --check <changed-python-files>
git diff --check

python -m pytest \
  tests/unit/models/test_lingbot_video_dit.py \
  tests/unit/models/test_lingbot_video_moe.py \
  tests/unit/pipelines/lingbot_video \

Validated GPU scenarios:

- 4x H100: Dense 1.3B SP4, 832x480, 121 frames
- 4x H100: MoE 30B + refiner, 1920x1088, 121 frames
- 4x H100: Dense CFG2×SP2, 832x480, 25 frames
- 4x H100: MoE FP8 CFG2×SP2 with co-resident refiner, 1920x1088, 25 frames
- H100: grouped_mm versus sorted eager MoE exact-parity check, 0 mismatches
- 4x H100: SP4 versus CFG2×SP2 Dense and MoE grouped_mm benchmarks

## Checklist

- [x] Code follows the project's coding standards (ruff)
- [x] Pre-commit hooks pass
- [x] Targeted LingBot-Video tests pass
- [x] New tests added for new functionality
- [x] Documentation updated (README, CLAUDE.md, docstrings)
- [x] Commit messages are clear and descriptive
- [x] PR title follows the convention: [TYPE] Brief description

## Related Issues

Fixes #

## Additional Notes

- expert_backend=fp8 is memory-oriented and requires PyTorch support for
  torch._scaled_mm; grouped GEMM remains the default four-GPU MoE backend.

- MoE expert parallelism and external FlashAttention are intentionally not enabled.
- The source-equivalent sorted eager MoE path remains the single-GPU correctness fallback.
- LingBot-Video usage documentation lives in examples/lingbot_video/README.md,
  consistent with other model-specific examples.

## GPU Architecture Support

- [ ] SM80 (Ampere, Ada Lovelace)
- [x] SM90 (Hopper H100)
  four-GPU refinement and releases both stages before high-resolution VAE decode.

- FP8 expert execution reduces expert-weight memory, but is currently slower than
  grouped BF16 GEMM and is therefore opt-in.