Prevent training service tests from leaking running jobs

[bghira]

Summary

• isolate training service unit tests from the persistent job repository so mocked start_training_job calls do not leave local jobs running
• reconcile all running local jobs by default on startup, retaining the optional cap for explicit callers
• add regression coverage for more than the old startup reconciliation cap

Validation

• .venv/bin/python -m unittest tests.test_training_service tests.test_local_gpu_allocator -v -f
• SIMPLETUNER_SELENIUM_TESTS=1 SELENIUM_SERVER_START_TIMEOUT=180 .venv/bin/python -m unittest tests.test_webui_e2e.TrainingMetricsDashboardTestCase.test_training_run_metrics_and_validation_media -v -f
• .venv/bin/python -m py_compile tests/test_training_service.py tests/test_local_gpu_allocator.py simpletuner/simpletuner_sdk/server/services/local_gpu_allocator.py
• git diff --check