feat(abot): add multi-session ABot-World serving and observability

[youngmagician114514]

Description

Adds ABot-World multi-session interactive serving with bounded process and
process-NCCL workers, deadline-aware continuation batching, opt-in CUDA Graph
execution, and reproducible serving observability tooling.

Motivation

ABot-World previously lacked an end-to-end runtime for multiple concurrent
interactive sessions. This PR adds explicit session ownership, lifecycle
cleanup, capacity enforcement, frame-credit-aware EDF scheduling, and raw
runtime facts for reproducible evaluation.

Type of Change

• Bug fix
• New feature
• Breaking change
• Performance improvement
• Code refactoring
• Documentation update
• Other

Changes Made

• Add multi-session ABot-World serving with process and process-nccl.
• Add session migration, bounded capacity, and EOS cleanup.
• Add frame-credit-aware EDF B1/B2/B3 batching and batch-time profiles.
• Add opt-in CUDA Graph with eager fallback and parity validation.
• Add trace replay, metrics, analysis tools, and monitoring configuration.

Testing

• Focused unit tests pass
• Manual multi-GPU serving validation performed
• Benchmarks and trace-validation tooling added/updated

Test commands:

PYTHONPATH=$PWD /public/fanyk1/lwb/envs/telefuser_sage291/bin/python -m pytest -q \
  tests/unit/pipelines/abot_world \
  tests/unit/service/livekit \
  tests/unit/validation

ruff check <changed-python-files>
ruff format --check <changed-python-files>
git diff --check
bash -n tools/validation/run_abot_4gpu_30min_trace.sh

Focused result: 138 passed.

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
• Commit messages are clear and descriptive
• PR title follows the convention: [Feature] Brief description

Related Issues

None.

Additional Notes

• CUDA Graph is opt-in and disabled by default.
• EDF batching is opt-in and preserves conservative default behavior.
• The supplied TurboServe trace is a public demo workload, not a released
  private production trace.
• Local results/, .orig, archives, and raw experiment files are excluded.

GPU Architecture Support

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

Performance Impact

The batching and CUDA Graph paths are opt-in and retain eager fallback.
Batch-time priors are profile-specific; this PR does not claim a universal
throughput improvement.