Offload debug rollout dump to a Ray task; drop routed_experts, add --save-rollout-interval

[nightlessbaron]

--save-debug-rollout-data serialized the full sample list with torch.save on the RolloutManager MainThread. With routed_experts in every sample the dump is ~200GB: ~30 minutes holding the GIL (frozen event loop, idle engines, blocked training handoff) and a peak-memory doubling that can OOM the head node.

Three changes:

• The write runs in a @ray.remote(num_cpus=1) task — its own worker process, so serialization never blocks the actor. At most one write is in flight; an overlapping dump is skipped with a warning instead of queued.
• rollout_routed_experts is dropped from the dump (it dominates the size and the training path already carries it), shrinking the payload ~40x and keeping the object-store handoff cheap.
• New --save-rollout-interval N dumps only every Nth training rollout (default 1; eval dumps unaffected).

Known trade-offs: the final rollout's dump is fire-and-forget, so a job that exits immediately after its last step can leave that one file incomplete; retention's walk-backward delete pairs with interval=1 (gaps from larger intervals can stop it early).