Fix agent daemon session process lifecycle

[RyanLee-Dev]

Summary

• add a shared CLI process runner that reliably waits for agent child processes and escalates cancel from SIGTERM to SIGKILL
• persist upstream engine sessions for agent-daemon runs so follow-up turns resume Claude Code, Codex, and Pi instead of starting detached histories
• scope daemon-side state by conversation/agent/engine, including stable Codex CODEX_HOME and Pi --session-dir
• avoid persisting failed Pi sessions and add process/session regression tests

Verification

• go test -count=1 ./apps/parsar-daemon/internal/agent/claudecode ./apps/parsar-daemon/internal/agent/codex ./apps/parsar-daemon/internal/agent/pi ./apps/parsar-daemon/internal/agent/clirunner ./server/internal/agentdaemon/binding ./server/internal/connector/agentdaemon
• make check
• rebuilt and restarted local Docker compose with clean volumes
• manually verified Claude Code, Codex, and Pi runs complete with no lingering claude, codex, or pi processes in the sandbox container
• manually verified Codex and Pi session rows are saved in agent_engine_sessions and resumed on subsequent turns