(fix) retry transient Ray ActorUnavailableError during rollout engine bringup

[EazyReal]

Problem

During colocated rollout-engine startup, the SGLang engine actors can already be healthy while Ray briefly reports a peer actor as unavailable because of a transient control-plane heartbeat / gRPC UNAVAILABLE miss:

ray.exceptions.ActorUnavailableError: ... RpcError ... UNAVAILABLE

Current main starts rollout engines first and then waits for the collected engine.init refs in RolloutManager.__init__. Encoder-disaggregated startup also has a synchronous encoder wait because encoder URLs are needed before language workers can start. A transient ActorUnavailableError at either wait boundary can abort startup even though re-awaiting the same already-submitted refs is safe.

Fix

Add one shared bounded retry primitive:

• slime.utils.retry.retry_with_backoff(thunk, *, should_retry, what, ...)
• The helper owns retry count, backoff, jitter, logging, and exhaustion behavior.
• The caller owns the retry predicate, so backend-specific classification stays at the call site.

Use that helper for rollout engine bringup waits:

• RolloutManager.__init__ waits on deferred rollout init handles through the helper.
• Encoder-disaggregated startup waits on encoder init handles through the helper.
• Only ray.exceptions.ActorUnavailableError is retried.
• Real init failures such as CUDA OOM/config/assert errors and permanent actor death still propagate immediately.

The important invariant is: retry the wait, not engine creation. The same ObjectRefs are re-awaited; no engine actors are recreated.

Consistency with follow-up retry PRs

This establishes the retry contract used by the onload follow-up: every bounded retry uses retry_with_backoff; rollout/Ray call sites only provide an idempotent thunk plus _is_transient_ray_unavailable.

Tests

CPU-only tests:

• tests/test_retry.py covers the generic helper: success after transient failures, exhaustion, immediate propagation of predicate-rejected errors, and not catching KeyboardInterrupt / SystemExit.
• tests/test_rollout_bringup_retry.py covers the rollout call site: transient actor-unavailable is retried on the same handles, non-transient errors are not retried, and the encoder wait path uses the same helper.

Validation

• PYTHONPATH=. uv run --with pytest pytest tests/test_retry.py tests/test_rollout_bringup_retry.py
• uv run --with pre-commit pre-commit run --files .github/workflows/pr-test.yml .github/workflows/pr-test.yml.j2 slime/ray/rollout.py slime/utils/retry.py tests/test_retry.py tests/test_rollout_bringup_retry.py