test: stabilize Windows CI without longer timeouts

[rekram1-node]

What

• cap unit-test Turbo concurrency at 3 on the 4-vCPU CI runners
• cancel service-smoke deadline timers when processes exit
• re-signal the compiled plugin watcher with valid source while retaining the existing 10-second failure bound

Why

#50449 exposed two separate Windows CI problems that should not be addressed by increasing test timeouts.

Turbo defaults to 10 concurrent tasks. That oversubscribes the 4-vCPU runners while each Bun test process performs its own work, stretching codemode tests from milliseconds locally into multi-second CI runs and tripping both Bun's default timeout and an existing performance assertion.

The compiled-service smoke also failed twice on v2 after a single plugin creation event was coalesced. It only polled afterward, so waiting longer could not recover. Rewriting valid plugin source periodically preserves the native-watcher coverage and gives the watcher another event without extending the deadline.

The smoke test also left losing 10-second Bun.sleep timers alive after each process exited. Clearing those timers reduced a local smoke run from about 11.5 seconds to about 1.4 seconds.

No test timeout is increased and the codemode performance assertion remains unchanged.

Testing

• bun test test/errors.test.ts test/limits.test.ts from packages/codemode
• bun typecheck from packages/cli
• compiled Bun service smoke, including 10 consecutive runs
• forced watcher retry path with the initial plugin write omitted
• compiled Node service smoke
• bun run check