[codex] Add sparse filesystem weight broadcast

[samsja]

Summary

Adds sparse checkpoint-format transfer as an opt-in mode of the existing filesystem weight broadcast backend.

• Adds shared, trainer, orchestrator, and inference config support for weight_broadcast.type = "filesystem" with weight_broadcast.sparse = true.
• Implements a trainer backend that writes full checkpoints on first/forced syncs and layerwise sparse delta directories otherwise, without gathering the full model state dict on rank 0.
• Adds an inference worker that materializes sparse deltas into a private local checkpoint cache before reusing the existing vLLM checkpoint reload path.
• Logs sparse broadcast metrics from the trainer backend; orchestrator W&B logging emits only sparse_broadcast_ratio so the main run shows the delta/full-weight size ratio directly.
• Preserves full base checkpoints needed by surviving sparse delta chains during broadcast cleanup.
• Keeps LoRA rejected for sparse filesystem broadcast until adapter semantics are defined.

Validation

• CLI parse check against examples/reverse_text/rl.toml with --weight-broadcast.type filesystem --weight-broadcast.sparse true --weight-broadcast.full-sync-interval 10
• uv run ruff check packages/prime-rl-configs/src/prime_rl/configs/trainer.py packages/prime-rl-configs/src/prime_rl/configs/orchestrator.py packages/prime-rl-configs/src/prime_rl/configs/inference.py packages/prime-rl-configs/src/prime_rl/configs/rl.py packages/prime-rl-configs/src/prime_rl/utils/validation.py src/prime_rl/trainer/rl/broadcast/__init__.py src/prime_rl/trainer/rl/broadcast/sparse_filesystem.py src/prime_rl/trainer/rl/train.py src/prime_rl/inference/vllm/server.py tests/unit/test_configs.py
• uv run pytest tests/unit/test_configs.py tests/unit/utils/test_sparse_weights.py tests/unit/orchestrator/test_scheduler.py -q
• uv run ruff check src/prime_rl/orchestrator/scheduler.py tests/unit/orchestrator/test_scheduler.py
• uv run pytest tests/unit/orchestrator/test_scheduler.py tests/unit/utils/test_sparse_weights.py -q
• uv run ruff check src/prime_rl/utils/sparse_weights.py src/prime_rl/trainer/rl/broadcast/base.py src/prime_rl/trainer/rl/broadcast/sparse_filesystem.py src/prime_rl/trainer/rl/train.py src/prime_rl/orchestrator/scheduler.py tests/unit/utils/test_sparse_weights.py
• uv run ruff check src/prime_rl/utils/sparse_weights.py src/prime_rl/trainer/rl/broadcast/sparse_filesystem.py src/prime_rl/inference/vllm/worker/sparse_filesystem.py src/prime_rl/trainer/rl/train.py packages/prime-rl-configs/src/prime_rl/configs/{trainer,orchestrator,inference,rl}.py tests/unit/utils/test_sparse_weights.py tests/unit/test_configs.py
• uv run pytest tests/unit/utils/test_sparse_weights.py tests/unit/test_configs.py -q