feat(rollouts) external rollouts endpoint with publish-only weight sync

[jvmncs]

Motivation

slime currently assumes it owns the rollout backend: it launches or registers SGLang engines (or fixed --rollout-external-engine-addrs) and pushes weight updates to each engine via per-engine RPCs. This PR lets slime train against an elastic, externally managed inference fleet behind a single HTTP endpoint — one with no stable per-engine handles and no SGLang-router worker-management APIs, where workers may scale up/down mid-run. Three opt-in pieces compose to support this:

1. Opaque HTTP rollout endpoint — send /generate requests to a base URL without launching or registering any SGLang workers.
2. Version-pinned rollout requests — generation payloads can pin an exact weight version, so an elastic fleet only serves samples from the intended policy version.
3. Publish-only disk delta sync — instead of direct update_weights_from_disk RPCs, the trainer publishes each complete delta version through a custom hook (e.g. to shared storage that the endpoint's workers consume), and the publish overlaps the next training step.

All features are off by default; existing behavior is unchanged when the new flags are unset.

Modifications

Opaque HTTP rollout endpoint (--rollout-http-endpoint-url)

• New slime/backends/sglang_utils/http_endpoint.py: URL normalization/validation and HttpEndpointRolloutServer, a no-engine rollout server stub (no offload/onload, no fault-tolerance recover).
• slime/ray/rollout.py returns it from the server-startup path; get_model_url() in sglang_rollout.py returns the endpoint URL (with the requested route appended) and never assumes router APIs exist.
• slime/ray/placement_group.py allocates no rollout GPUs in this mode.
• Mutually exclusive with --rollout-external-engine-addrs (validated).
• --rollout-http-endpoint-abort-strategy {cancel-only,router-workers}: cancel-only (the default when an endpoint is set) cancels local pending tasks without calling router /list_workers; the existing router-based abort is refactored into _drain_aborted_pending_tasks and remains the default otherwise.

Version-pinned rollout requests (--rollout-weight-version-policy exact-rollout-id)

• /generate payloads gain weight_version={"exact_version": rollout_id}, scoped per rollout via rollout_weight_version_context.
• Retries while the target version is unavailable are tunable via --rollout-weight-version-retry-attempts/-sleep; slime/utils/http_utils.py post() gains a retry_sleep parameter.

Publish-only disk delta sync (--update-weight-delta-publish-only), and skips version-dir cleanup (--update-weight-delta-keep-files is required).

• The dispatched publish intentionally stays in flight across the training step and is drained at the start of the next sync (or on disconnect_rollout_engines), so publish latency overlaps training; a failed publish surfaces one sync late on rank 0.
• Argument validation enforces delta mode + disk transport + publish path + keep-files.

Tests

• New: tests/test_rollout_http_endpoint.py (8 tests: URL validation, endpoint routing, payload pinning, retry-until-version-available, cancel-only abort, no-engine server), tests/test_delta_publish_only.py (4 tests: hook invocation without engine RPCs or cleanup, no-op version publish, drain-on-disconnect, publish deferred to finalize).
• Extended: tests/test_placement_group.py (http-endpoint layouts), tests/test_megatron_argument_validation.py.

Checklist

• Format your code
• Add unit tests according — 12 new tests + 3 extended parametrizations, all passing
• Update documentation
• Provide accuracy and speed benchmark results