refactor: speed up unit tests and trim redundancy

[xavierlyu]

Summary

• Mock _tokenize in the run_baseline test to remove a ~300s accidental network wait
• Remove ~44 redundant unit tests (dead schema, constant assertions, duplicate scenarios, low-value defensive cases)
• Add pytestmark = pytest.mark.unit to GPU orchestration and cacheon_db loaders tests so CI runs them

---

Note

Low Risk
Changes are confined to the test suite; no production behavior is modified, only coverage breadth is reduced in a few edge-case areas.

Overview
Test-only PR that shortens CI runtime and reduces duplicate or low-value coverage; production code is unchanged.

The largest fix is in test_docker_eval: run_baseline error-path tests now @patch validator.docker_eval._tokenize, avoiding real HTTP to vLLM’s /tokenize (previously could hang for minutes).

Several modules shed redundant cases: BaselinePromptResult / BaselineCache round-trips in test_baseline; EvaluationJob / EvaluationResult and many schema default tests in test_validator_eval_schema; constant and alias assertions in test_prompts; duplicate digest and _effective_overtake_threshold tests in test_validator_state; “never raises” Docker/GPU cleanup tests in test_docker_eval; eval-progress no-file / bad-index / OSError swallow tests in test_eval_progress; and similar overlap elsewhere. test_cpu_validator merges dry-run vs live weight-setting into one @pytest.mark.parametrize test.

pytestmark = pytest.mark.unit is added to GPU orchestration, GPU session, GPU SSH, and test_cacheon_db_loaders so those files run in the unit CI job.

^{Reviewed by Cursor Bugbot for commit 241dc37. Bugbot is set up for automated code reviews on this repo. Configure here.}