[CI] add reward-fn cpu tests + cpu-on-main trigger (port slime #1939+#1940)

[aoshen02]

Ports THUDM/slime#1939 (add more cpu ci) + #1940 (run cpu test on main).

What

• tests/test_rm_{deepscaler,f1,gpqa,math,math_dapo}.py — pure-cpu unit tests for vime/rollout/rm_hub reward fns. 107 tests, validated locally (all pass).
• vime/utils/misc.py — lazy import ray so cpu-only paths don't require ray.
• pr-test.yml.j2 — register the 5 rm tests in the cpu (num_gpus:0) matrix; enable push-to-main trigger gated to cpu jobs only (GPU/self-hosted jobs stay PR/label-gated, so push never burns the fleet).

Port notes (sglang→vllm / vime divergence)

• slime #1939 also added tests/test_sample.py, which references Sample.rollout_id (introduced by the #1926-chain → vime [Train] Micro-batch scheduling on rollout side + variable global batch size (port slime #1926+#1930+#1933) #119). That test is stacked on [Train] Micro-batch scheduling on rollout side + variable global batch size (port slime #1926+#1930+#1933) #119, not here.
• test_sample.py's arg fixture used sglang_speculative_algorithm; vime's Sample reads vllm_speculative_config — translated on the [Train] Micro-batch scheduling on rollout side + variable global batch size (port slime #1926+#1930+#1933) #119 branch.

Validation

• Local: pytest tests/test_rm_*.py → 107 passed.
• ruff clean. CI (cpu job) will run the rm tests on this PR.

Part of the #107 slime→vime sync. 🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request adds comprehensive CPU unit tests for several reward model scoring modules, including deepscaler, f1, gpqa, math_utils, and math_dapo_utils. It also updates vime/utils/misc.py to lazily import ray inside get_current_node_ip(), allowing CPU-only paths to run without requiring the ray library. The review feedback correctly identifies two issues in the new tests where integer values were used instead of floats, which prevents the float coercion and parsing logic from being properly exercised in test_rm_deepscaler.py and test_rm_math_dapo.py.

[gemini-code-assist]

The test is intended to verify float labels, but it passes the integer 42 instead of the float 42.0. This makes it identical to test_label_as_int_is_coerced_to_string and fails to actually test float coercion.

Suggested change

	assert get_deepscaler_rule_based_reward(r"</think>\boxed{42}", 42) == 1
	assert get_deepscaler_rule_based_reward(r"</think>\boxed{42}", 42.0) == 1

[gemini-code-assist]

The test docstring mentions that floats collapse to canonical int strings, but the test currently passes "42" as the ground truth, which does not exercise the float parsing path. Passing "42.0" instead would properly validate this behavior.

Suggested change

	correct, pred = is_correct_minerva("Long solution. Answer: 42", "42")
	correct, pred = is_correct_minerva("Long solution. Answer: 42", "42.0")

[aoshen02]

Closing for now (slime→vime sync being restructured). Branch is preserved — can reopen anytime.