[claude]

Add MatbenchTask.unload() to free the dataset from memory and keep recorded results.

• unload() sets task.df = None, then runs gc.collect().
• If the task is not loaded, it does nothing.
• Added test: test_unload in matbench/tests/test_task.py.

Reviewer checks

• Only adds a method. Results, scores and defaults do not change.
• record() still needs the dataset. Load it again before recording more folds.
• No MatbenchBenchmark.unload(). Add one in a later PR if you want it.
• The issue's other question (other large objects to delete) is not handled here.

Fixes #197

[claude]

CI test (3.9) fails, so this can't merge yet.

• I couldn't open the job log from here.
• Fix #162: Tasks with uncertainty should enforce all folds have uncertainty #392, on the same base, passes test (3.9), so the failure is probably in this PR.
• The new test_unload is the most likely cause. Fix #134: Trying to load a partially recorded benchmark from disk raises ValueError #394 fails in the same way.
• @ardunn: please check the log or re-run it. If it is a real failure, comment @claude fix <cause>.

[liuxiaotong15]

LGTM. All checks pass.