Skip unparseable gold solutions in cosine_scaled_reward

[shaurya416]

Fixes #732

Problem

get_cosine_scaled_reward returns 1.0 for a completion whenever the gold solution cannot be parsed (rewards.append(1.0) # Skip unparseable examples). 1.0 is the default max_value_correct, so an example with a malformed or missing gold answer gives the model the maximum reward for any output, including one that is plainly wrong. The comment says the example is skipped, but a numeric reward is not a skip.

accuracy_reward handles the same case by returning None, so the trainer skips the example (#566). This change does the same in cosine_scaled_reward.

Change

• rewards.append(None) instead of 1.0 when len(gold_parsed) == 0, and the return annotation list[Optional[float]] as on accuracy_reward.
• Two tests in tests/test_rewards.py: an unparseable gold gives [None, None] for completions of different lengths, and in a mixed batch only the unparseable example is None.

Not changed: len_reward has the same shape (correctness.append(True)), but test_unparseable_solution pins its current 0.5, so I left it for a separate decision. The optional follow-up from the issue (dropping extraction_config=[LatexExtractionConfig()] on the gold parse so bare answers stay scorable) is also left out.

Validation

I ran the real tests/test_rewards.py against the real rewards.py with math-verify==0.5.2 and latex2sympy2_extended, from a stub package that replaces only the modules the reward functions under test never call (code_providers, competitive_programming, configs). TestGetRewardFuncs was deselected because it needs the real configs.

	new tests (2)	existing tests
main	2 failed	45 passed, 1 failed
this change	2 passed	45 passed, 1 failed

The one failing existing test, TestRepetitionPenaltyReward::test_full_repetition_with_language, fails the same way before and after because transformers is not installed in this environment. ruff format --check --line-length 119 passes on the two files.