perf(test): reuse one Postgres server for integration databases

[mforce]

Change

• Start one pinned Postgres Testcontainer per integration test process. Clone a migrated template into a distinct database for each factory; hand virgin databases to migration and startup-disabled tests.
• Preserve each race's actors in one database and the dedicated one-slot app pool test.
• Scope the first-run advisory-lock assertion to its database and add a held-lock positive control. The diagnosis and all configurations are in docs/plans/863-integration-collection-split/02-shared-postgres-spike.md.

Measurement

• Local baseline, three full suites: 388.7, 388.6, 384.4 s; 128 Postgres containers, 138 total.
• One-server path, three full suites: 344.1, 343.8, 346.7 s; one Postgres container, 11 total. Median wall improvement: 44.5 s (11.5%). The final staged source passed a further full run in 337.6 s.
• CI integration Test steps on the code commit, three attempts: 535, 487, 424 s (median 487). The final documentation-only head added a 360 s pass on the same test code. The four prior fix(test): repair demo seed false pass and measure collection floor #1000 steps were 605, 606, 390, 596 s (median about 601). The ranges overlap, so CI supports a likely gain without a precise saving estimate. All four new runs passed 1,873 tests and created 11 containers.
• The DemoSeed class split was measured separately: B-only 393.5, 368.6, 387.3 s; A+B 352.7, 347.5, 350.9 s. It did not lower the suite wall, so this PR leaves the class together.

Correctness and checks

• Real-Postgres probe: the worker's exact (271, 1) advisory key conflicts in the same database and succeeds in another, in both directions. The first-run lock query sees a same-database held key; a temporary mutation made its zero-lock assertion fail Expected: 0; Actual: 1. The restored test passes and retains a positive control.
• The server defaulted to 100 connections; the sampled peak was 39 clients across 13 active databases. Twenty parallel template clones completed without failure. The virgin-database, leader-lease, one-slot-pool, and intermediate-migration focused tests passed.
• Full Release integration suite: 1,873 passed on final code. The tracked-file Postgres image-pin guard passed after the final documentation edit. dotnet build passed with zero warnings.

Refs #863.