Sub-agent: auto-cancel hung agents to prevent slot exhaustion

[Hmbown]

Stuck sub-agents remain in Running status indefinitely, consuming slots from the max_agents limit (default 10, hard max 20). No global heartbeat or timeout auto-cancels hung sub-agents, and the cleanup() method only removes terminal agents older than 1 hour. This causes #2603 — users can't start new sessions because slots are full of zombie sub-agents. Fix: add a configurable sub-agent heartbeat timeout (default 5 minutes) that auto-cancels non-responsive agents.

Source: investigation of #2603 slot exhaustion from hung sub-agents.

[Hmbown]

Implemented on the v0.8.52 stabilization branch. Running sub-agents now track manager-visible progress and auto-cancel after [subagents] heartbeat_timeout_secs with a 300s default, releasing the slot instead of leaving the parent turn stuck forever.

Validation so far: targeted heartbeat/cleanup tests, cargo test --workspace --all-features --locked, cargo clippy --workspace --all-targets --all-features --locked -- -D warnings, npm wrapper tests, release version drift checks, and crates dry-run all pass.