feat(inference): allow running a policy ref that isn't tied to a LeLab job

[ravediamond]

Closes #111.

lerobot already supports third-party policy plugins (a pip-installed lerobot_policy_<name> package registers into PreTrainedConfig's choice registry, and lerobot-rollout already calls register_third_party_plugins() at startup) — and _resolve_policy_path in lelab/rollout.py already accepts an arbitrary local checkpoint dir or Hub ref. The actual gap was the UI: the inference modal only ever built policy_ref from a job's own checkpoints, so there was no way to point at a checkpoint that isn't tied to a LeLab training run.

Added a "Use a custom policy path or ref" toggle to the inference modal. When on, you type a local checkpoint directory or a Hub ref (same user/repo@root / user/repo@checkpoints/000050 shape as job checkpoints) instead of picking from the dropdown.

New GET /policy-config?policy_ref= endpoint reads that ref's config.json and returns the same summary (policy type, expected cameras, whether it needs a task string) the job-checkpoint route already returns — pulled the shared extraction into jobs.summarize_policy_config() so both routes use the same logic.

Known limitation, not fixed here: requires_task is driven by a hardcoded list of language-conditioned policy types (smolvla, pi0, pi0_fast, pi05). A plugin policy that needs a task string won't get the task input shown automatically — you can still type one, the field just won't appear on its own. Didn't want to guess at a convention for plugins to declare this.

Tested: pytest (273 passed, 6 new tests), ruff check/format, tsc, eslint, prod build all clean. Not tested against a real third-party plugin package — don't have one installed to verify with.