feat(checkpoint): support sharded Gym turn recovery

[macandro96]

Summary

• coordinate turn-level Gym checkpoint prepare, commit, restore, resume, and abort across multiple shards
• persist shard-local artifacts under deterministic gym-shards/<instance>/ paths
• merge shard topology while deduplicating the shared policy-model participant
• route recovered continuations and completion ACKs back to their owning shard
• retain the one-replica-per-shard guard

Stack

• depends on perf(checkpoint): reduce Gym control-plane overhead #4321
• next: sharded generation-prefix recovery

Validation

• ruff check and ruff format --check passed on the sharded-turn tree
• Python compilation passed
• the two-shard turn-recovery functional test still needs to run on the cluster

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.