Add Megatron-Bridge LoRA support for GRPO actor training

[taivu1998]

Summary

Addresses #1202.

This PR adds a first supported Megatron-Bridge LoRA path for dense GRPO actor training in slime. It introduces LoRA CLI flags, validates the initially supported configuration at startup, applies Megatron-Bridge PEFT LoRA only to the actor model, and exports effective actor weights to SGLang by temporarily merging adapters into the live model during bridge-based HF weight conversion.

Why

Issue #1202 asks for LoRA support for GRPO training and examples. The discussion also calls out known Megatron-Bridge LoRA risk around MoE and checkpointing paths, so this implementation intentionally starts with a narrow, guarded dense-model path rather than silently enabling unsupported combinations.

Changes

• Added --enable-lora, --lora-target-modules, --lora-rank, --lora-alpha, and --lora-dropout.
• Added validation for the first supported LoRA slice:
  • Megatron backend.
  • Megatron-Bridge HF export mode.
  • GRPO actor training.
  • Colocated rollout outside debug train-only runs.
  • Dense models only.
  • Default weight backuper enabled.
• Added a Megatron LoRA helper module that:
  • lazily imports Megatron-Bridge PEFT LoRA.
  • builds the LoRA config from slime args.
  • applies LoRA only to the actor provider path.
  • logs local trainable and total parameter counts.
  • restores TensorBackuper-style model weights around temporary LoRA merge export.
• Updated bridge weight export so LoRA runs export base + adapter effective actor weights to SGLang, then restore unmerged training weights.
• Added focused unit tests for validation, config mapping, actor-only application, merge traversal, and backup restore safety.
• Added an English advanced usage page with a concrete GRPO LoRA flag example and linked it from the docs index.

Guardrails

The PR rejects combinations that need separate parity work before support:

• MoE models.
• PPO or critic-based training.
• Decoupled rollout mode outside --debug-train-only.
• Custom model providers.
• --only-train-params-name-list and --freeze-params-name-list.
• On-policy distillation.
• Reference model update intervals.
• --disable-weights-backuper.

Validation

• env UV_CACHE_DIR=/tmp/uv-cache PYTHONPATH=. uv run pytest tests/test_lora_support.py -> 22 passed
• python3 -m py_compile slime/backends/megatron_utils/peft.py slime/backends/megatron_utils/model_provider.py slime/backends/megatron_utils/model.py slime/backends/megatron_utils/update_weight/hf_weight_iterator_bridge.py slime/utils/arguments.py tests/test_lora_support.py
• git diff --check

uv run ruff check ... was attempted locally but could not run because this worktree environment does not have a ruff executable installed.