feat: add Tinker-compatible training API

[yushengsu-thu]

Summary

• add a self-hosted data plane compatible with the official Tinker SDK ServiceClient, TrainingClient, and SamplingClient routes
• back forward/backward, optimizer steps, exact training-state checkpoints, and retained gradients with Megatron multi-LoRA
• expose immutable named and ephemeral SGLang LoRA sampling snapshots, base-model sampling, auth, readiness, docs, and a runnable SDK smoke test

Compatibility

• supports encoded-text LoRA training, CE/importance-sampling/PPO/CISPO/DRO/custom losses, top-k logprobs, Adam, module groups, seeds, unload, telemetry, and weights info
• supports DP and TP; validates and rejects unsupported PP, CP, BSHD, colocation, full fine-tuning, and non-LoRA modes at startup
• implements the official SDK core data plane, not hosted account/project/billing/admin control-plane APIs

Validation

• pytest -q tests/fast: 4081 passed, 22 skipped; one Ray dashboard test collided with an already-running external Ray cluster and passed in isolation after stopping that cluster
• targeted final regression suite: 96 passed
• official Tinker SDK 0.24.0 end-to-end smoke passed with DP=2 and TP=2, including exact optimizer/retained-gradient restore and all sampling modes
• Ruff format/lint, compileall, shell syntax, docs JSON, and git diff --check passed

Tested dependency heads

• Miles main: 2cb59235176a170480765e762e85a952e6ba72b7
• SGLang sglang-miles: 3003d70f680d41c59d1b7acbf65cb47795dfd19e
• Megatron-Bridge bridge: 7f0fb3456f8ffe47599b5fd167b454605d85f932

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.