feat(adapter): add Megatron Bridge and NeMo RL checkpoint adapters

[g-husam]

Megatron Bridge exposes a CheckpointManager protocol that a run can swap in via CheckpointConfig.custom_manager_class. MLFlashpointBridgeCheckpointManager implements it: non-persistent checkpoints (non_persistent_save_interval) are written by ML Flashpoint into node-local memory, while durable checkpoints (save_interval) keep going through Megatron Bridge untouched. Bridge takes the non-persistent branch only on steps that are not also durable-checkpoint steps, so the two cadences never collide.

Resume works by publishing an MLFlashpointLocalCheckpointIndex under checkpointing_context["local_checkpoint_manager"], which is what Bridge's setup._should_load_checkpoint consults. Load prefers the newest recoverable ML Flashpoint container and falls back to Bridge's own path when there is none or when the in-memory read fails.

NeMo RL needs a different entry point: it builds its Megatron state with Megatron Bridge but calls Bridge's functional save_checkpoint directly from MegatronPolicyWorker, so custom_manager_class is never consulted anywhere in that repo. install_into_worker wraps that one method instead, in either "augment" mode (durable writes unchanged, ML Flashpoint alongside) or "replace" mode (only every Nth checkpoint stays durable).

Also adds a checkpoint-timing benchmark harness that parses Megatron's save-checkpoint timers and ML Flashpoint's own execution timings out of training logs and diffs two runs, plus a runbook for measuring the difference on a GKE training cluster, and source-verified research notes on all three codebases for future work.