feat(experiment): add single-episode orchestrator

[Staaaaaaaaar]

Summary

• add the rh_experiment ROS 2 package and a ROS-independent single-Episode lifecycle controller
• load and statically validate an Experiment config, explicitly rejecting multi-Episode configs in the PR 07 scope
• wait for fresh Env and Agent readiness, then enforce Env reset before Agent reset using bounded steady-time service deadlines
• publish the authoritative Episode state and support manual/automatic start, operator abort, component failure handling, and one-time termination commit
• serialize every lifecycle mutation on one control thread and cancel unclaimed control requests that exceed their response deadline
• perform a bounded safe-stop delivery handshake before publishing FINISHED

Runtime semantics

• lifecycle: PREPARING -> READY -> RUNNING -> TERMINATING -> FINISHED
• /roboharness/episode/state has one authoritative publisher
• component ERROR, stale heartbeat, reset rejection, or reset timeout commits the corresponding infrastructure termination reason
• manual and automatic execution share the same start transition
• termination reason is immutable after the first accepted candidate
• safe stop confirms reliable DDS/RMW delivery of TERMINATING to matched subscriptions; the current wire contract does not provide a stronger component-level semantic acknowledgement

Scope

PointNav task publication, evaluator conditions, result recording, Isaac integration, and multi-Episode iteration remain in later roadmap PRs. PR 08 will insert task publication after Env/Agent reset and before READY.

Validation

• make dev-check
• 11 ROS packages built
• 164 tests passed, 0 errors, 0 failures, 0 skipped
• ros2 pkg executables rh_experiment reports rh_experiment orchestrator
• integration coverage includes reset ordering, happy path, illegal start, automatic start, abort, stale status, reset timeout, safe-stop delivery, and one-time termination