Add topology-portable trainer checkpoints

[bradhilton]

Summary

• replace separate trainable/frozen LoRA slots with checkpoint-based TrainerRank state
• add exact canonical checkpoint serialization for LoRA + FP32 Adam state
• restore checkpoints transactionally across supported TP/PP/EP/ETP/CP/DP topologies
• pin and validate immutable base-model revisions through standard PEFT adapter metadata
• add task/context checkpoint loading and service-neutral LoRA export
• preserve logical checkpoint identities while allowing pinned local materialization
• add bounded checkpoint-save bookkeeping and checkpoint performance harnesses

Validation

• repository Ruff, format, ty, and lock hooks
• exact 200-test Megatron lightweight CI slice
• full 59-test canonical checkpoint codec suite
• 605 backend/tinker CI-slice tests (3 skipped)
• 1→2, 2→1, DP deduplication, PP, EP, ETP, and CP checkpoint tests
• next-optimizer-step numerical equivalence and transactional failure coverage
• GitHub quality, install smoke, and 2×H200 checks green on final commit

Companion

• Caladan: https://github.com/OpenPipe/caladan/pull/102

Remaining draft acceptance

The isolated cross-repo H200 acceptance/performance matrix and combined production-LOC constraint decision are still pending.