fix(neon_local): allow restart of crashed compute endpoints

[hawthornic]

Problem

When a compute endpoint is killed externally (kill -9), neon_local cannot restart it:

• cargo neon endpoint stop fails because pg_ctl cannot signal a dead PID
• cargo neon endpoint start fails because check_conflicting_endpoints sees the crashed endpoint as a "duplicate primary" (status is Crashed, not Stopped)
• User is stuck

Root cause

check_conflicting_endpoints iterates all endpoints and never excludes the endpoint being operated on. The status filter only excluded Stopped, not Crashed. Additionally, stop() unconditionally called pg_ctl stop which fails when the postmaster is already dead.

Fix

1. check_conflicting_endpoints: Added exclude_id: Option<&str> parameter so an endpoint does not conflict with itself when being restarted. The Create path passes None (no-op), the Start path passes the endpoint being started.

2. stop(): Before attempting pg_ctl stop, check if the endpoint is Crashed. If so, verify the postmaster process is truly dead via kill(pid, None) (signal 0 — existence check only), then clean up the stale postmaster.pid and let compute_ctl shut down. If the postmaster is alive despite the Crashed heuristic (300ms TCP timeout false-positive during startup), fall through to the normal stop path.

Testing

Manual testing confirmed all three paths work:

• kill -9 → direct cargo neon endpoint start ✅
• kill -9 → cargo neon endpoint stop → start ✅
• Normal stop / start unaffected ✅

A regression test in test_runner/ is planned but not included in this PR.