[Subtask] Run core database test lanes with bounded adaptive forks

[nevzheng]

Describe the subtask

Run the explicit H2, MySQL, and PostgreSQL core lanes with conservative Gradle worker-level concurrency after #13518 establishes class-scoped fixture ownership. The goal is to reduce lane wall time without enabling JUnit method-level parallelism or sharing database servers between workers.

The change should:

• derive maxParallelForks from CPU, JVM-visible memory, Gradle worker capacity, concurrently selected database lanes, and a conservative rollout cap;
• support strict property and environment overrides, enforce every known bound, and fall back to one worker when automatic resource signals are incomplete;
• retain one worker for macOS fixed-network Docker configurations;
• add exact-type DatabaseTestContext injection for the selected backend and isolation policy;
• preserve [#6505] improvement(test): Reuse core database fixtures by test class #13518’s cleanup, poisoning, rebuild, and fresh-fixture semantics within each worker;
• keep JUnit method-level parallelism disabled;
• emit per-worker and per-class [CORE-DB-SHARD] telemetry in each lane’s evidence artifact; and
• add focused policy, injection, and telemetry regression tests.

Acceptance criteria:

• Unit, H2, MySQL, and PostgreSQL lanes pass, and every database lane preserves the exact test identity/count baseline from [#6505] improvement(test): Reuse core database fixtures by test class #13518.
• Every Gradle test worker owns its database server; no server or connection pool is shared across workers.
• The selected worker count never exceeds any known CPU, memory, Gradle-worker, selected-lane, or rollout bound.
• Malformed, non-positive, overflowing, and above-bound overrides fail with actionable errors; incomplete automatic signals select one worker.
• The macOS fixed-network Docker guard selects one worker where applicable.
• Matched before/after runs report task duration, worker count, fixture initialization count, and container-start count.

Connection pooling, shared database servers, namespace sharding, and method-level parallelism are out of scope.

Depends on #13518. Related to #13517.

Parent issue

#13515