fix(agent): handle SIGTERM and detect orphaned sessions

[pando85]

Problem

When OpenCode restarts, the Playwright MCP server process (launched via passless agent playwright-mcp) becomes orphaned but keeps running. Passless tracks this as an active session and refuses to start a new one, returning "profile 'coding' already has an active session (pid XXXX)".

Root Causes

1. No SIGTERM handling in the client - Only SIGINT was handled via ctrlc::set_handler. When OpenCode sends SIGTERM, the client died without calling terminate_principal.
2. No client-liveness tracking - The daemon had no concept of which admin client requested the session, so it couldn't detect when the client died.
3. Hardcoded 1-hour session TTL - Ignored the profile's max_session_ttl config.
4. No PR_SET_PDEATHSIG on principal - The principal wouldn't receive a signal if the daemon died.

Changes

• Add SIGTERM signal handler in dispatch_run to properly revoke sessions when the client process receives SIGTERM
• Track requesting client PID in ManagedPrincipalSession and detect orphaned sessions in reap_expired_sessions by checking if client is alive via /proc/{pid}/stat
• Use profile's max_session_ttl instead of hardcoded 3600s for session deadline
• Add PR_SET_PDEATHSIG on principal process so it receives SIGTERM if daemon dies
• Add signal-hook dependency for portable signal handling

Testing

• All 1465 existing tests pass
• Build succeeds with --features agent
• Manual testing confirms orphaned sessions are now cleaned up when client dies