perf: fix N+1 queries in load_posts_keyed, add DB pool timeout

[ety001]

Summary

Three targeted performance fixes to address recurring hivemind outage (every 1-2 days, EB health goes Yellow with ~46% 4xx).

Root cause

get_discussion pipeline exhausts the 25-connection aiopg pool under burst traffic, starving health checks → ELB marks instance unhealthy → EB replaces it.

Changes

1. Eliminate N+1 community queries (objects.py:load_posts_keyed)

   • Merged per-community title + role queries into 2 batch queries
   • Reduces DB round-trips from 2N to 2 (N = number of communities in thread)
   • A thread spanning 30 communities previously issued 60 extra queries

2. Add connection pool acquisition timeout (db.py)

   • aiopg.sa.create_engine(timeout=10)
   • If all connections are busy, fail fast in 10s instead of blocking indefinitely
   • Normal queries complete in <1s; 10s is a safe upper bound
   • Prevents connection starvation cascading into health check failure

3. Increase _child_ids cache TTL (thread.py)

   • 30s → 120s
   • Reduces repeated BFS lookups on deep comment trees
   • Most comment trees are static after initial load

Verification

• make fmt ✓
• python3 -m py_compile ✓ (all 3 files)
• Logic reviewed against production Scalyr logs showing get_discussion as 50% of failure traffic