[REFACTOR] Replace in-tree cache_mem with CacheSeek integration

[yJader]

Co-authored-by: @yx0716

Description

This PR replaces TeleFuser's in-tree latent cache implementation with an optional CacheSeek integration path. It wires CacheSeek into the service container, task service, CLI flags, Wan2.2 service examples, and LingBot World Fast world-KV hooks while keeping latent cache disabled unless explicitly requested.

Motivation

Cross-request latent/KV reuse is now owned by CacheSeek instead of TeleFuser-local cache_mem code. TeleFuser should depend on CacheSeek only when the feature is enabled, fail clearly when CacheSeek is missing, and keep the default import/runtime path lightweight when latent cache is disabled.

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

• Removed the in-tree telefuser/cache_mem implementation and related unit tests.
• Added lazy CacheSeek service initialization in TeleFuser service container and task service code.
• Added/updated latent cache CLI and server config plumbing, including direct failure when latent cache is enabled but CacheSeek is unavailable.
• Updated Wan2.2 T2V service examples for CacheSeek-backed latent cache and added a nocache service example.
• Added LingBot World Fast world_kv_binding runtime hooks so CacheSeek exact-prefix reuse can fast-forward cached chunks.
• Updated English and Chinese latent cache docs with CacheSeek usage and the CacheSeek GitHub link: https://github.com/Tele-AI/CacheSeek.

Testing

• Unit tests pass (pytest tests/)
• Manual testing performed
• Benchmarks added/updated (if applicable)

Test commands:

# TeleFuser targeted latent-cache/service tests.
python -m pytest \
  tests/unit/service/test_latent_cache_cli.py \
  tests/unit/service/test_latent_cache_task_service.py \
  tests/unit/pipelines/wan_video/test_service_examples.py \
  tests/unit/pipelines/wan_video/test_latent_data_utils.py -q
# Result: 15 passed

# Real Wan2.2 service e2e without latent cache.
# Model: Wan2.2-T2V-A14B
# Config: num_inference_steps=2, num_frames=5, resolution=480p, parallelism=1.
# Result: completed; non-empty mp4 generated.

# Real Wan2.2 service e2e with CacheSeek enabled.
# Model: Wan2.2-T2V-A14B
# Cache mode: read_write
# Result: task_status=completed; non-empty mp4 generated.
# Cache evidence: audit log contains lookup_hit skip_step=1 and save_stored.

# Real LingBot World Fast exact-prefix e2e through TeleFuser world_kv hooks.
# Note: uses CacheSeek as an external dependency; this PR only includes TeleFuser-side hooks.
export CUDA_VISIBLE_DEVICES=0,1
export LINGBOT_WORLD_CHECKPOINT_DIR=<lingbot-world-fast-checkpoint-root>
export WORLDKV_REPO_ROOTS=<telefuser-repo>:<cacheseek-repo>
export PYTHONPATH=<telefuser-repo>:<cacheseek-repo>:${PYTHONPATH:-}
cd <cacheseek-repo>
python \
  examples/exact_prefix_reuse/e2e_telefuser_lingbot.py \
  --frame-num 13 \
  --prefix-chunks 1 \
  --out-dir <output-dir> \
  --image-path <lingbot-example-image> \
  --action-path <lingbot-example-action-dir> \
  --aux-device cuda:1 \
  --no-save-videos
# Result: all_pass=true; fast_forward_k A=0, B=1, C=1, D=0.

Additional validation notes:

• Wan2.2 CacheSeek audit log contained lookup_hit skip_step=1 and save_stored.
• LingBot e2e manifest reported all_pass=true.
• LingBot e2e log reported world_kv: fast-forward 1 chunks (decode-only).
• GPUs were checked after e2e runs and had no remaining compute processes.

Checklist

• Code follows the project's coding standards (ruff)
• Pre-commit hooks pass (pre-commit run --all-files)
• All tests pass (pytest tests/)
• New tests added for new functionality
• Documentation updated (README, CLAUDE.md, docstrings)
• Commit messages are clear and descriptive
• PR title follows the convention: [TYPE] Brief description

Related Issues

N/A

Additional Notes

This PR is scoped to the TeleFuser-side CacheSeek adaptation. It does not include CacheSeek repository changes. The LingBot e2e command above exercises CacheSeek as an external dependency to verify that the TeleFuser world_kv_binding hooks are usable end to end.

GPU Architecture Support

• SM80 (Ampere, Ada Lovelace)
• SM90 (Hopper H100)
• SM100+ (Blackwell)

No kernel-specific code was changed. Real e2e validation ran on NVIDIA H100.

Performance Impact

No kernel-level performance change is intended. CacheSeek reuse can reduce repeated work when enabled. The LingBot exact-prefix smoke e2e showed functional reuse:

• A cold run: 7.456s
• B full hit: 1.519s
• C prefix hit: 1.494s
• D cold fork reference: 4.628s

These are smoke e2e timings on H100 and should not be treated as a formal benchmark.