feat: Parquet format and rolling retention for rollout debug saves

[nightlessbaron]

Adds two new flags to --save-debug-rollout-data:

• --save-rollout-format {pt,parquet} — write snappy-compressed parquet instead of .pt; rollout_id stored in schema metadata. Default pt keeps existing behaviour.
• --save-rollout-retain-last-n N — after each training rollout, delete the file for step R-N. O(1), no directory scan. Default 0 keeps all files.

Eval rollouts (eval_*) are excluded from the retention window.

Diagnostics are written by @odp. We might remove them in the future depending on how compute intensive they are.

[odp]

@nightlessbaron in miles/ray/rollout.py

    def _get_rollout_data(self, rollout_id):
        if self.args.load_debug_rollout_data:
            data = torch.load(
                self.args.load_debug_rollout_data.format(rollout_id=rollout_id),
                weights_only=False,
            )["samples"]

the load_dubug_rollout_data path still expects a torch tensor. May be we don't do parquet in this PR just do --save-rollout-retain-last-n which helps save disk space. The parquet conversion can be done outside when we need it.

[nightlessbaron]

made _get_rollout_data dispatch on the .parquet extension so load and save stay consistent without an extra flag. also added the logger.info on delete and switched the single unlink to a backward walk so we drain all stale files when restarting with a smaller N (not just one).

[odp]

LGTM!