fix(workspace): back off unreachable database diagnostics

[psteinroe]

Prevent DB-backed workspace operations from retrying a known-unreachable database connection on every edit. Connection failures are cached per connection key in the connection manager with exponential backoff, and DB callers now route through the manager so failures are tracked consistently.

Diagnostics Resilience

Database-backed file diagnostics are best-effort. When the database is unreachable, parser and static diagnostics are still returned instead of failing the whole edit-triggered diagnostics request.

Connection Manager Routing

DB-backed diagnostics, schema-aware static analysis, splinter, completions, hover, statement execution, and schema-cache invalidation now use ConnectionManager::with_pool, so connection failure backoff is managed centrally.

Regression Coverage

Added a workspace test that points at a closed local Postgres port, verifies static diagnostics are still produced, and confirms the second diagnostics pull during backoff completes quickly.

[juleswritescode]

LGTM — one concern: we're still using get_current_connection() a whole bunch in server.rs, for example when we fetch diagnostics or completions. We should probably migrate that to with_pool as well