[evals] Add diff and patch PPL eval sets

[dlwh]

🤖 Part of #5005.

Description

Add PPL/gap eval sets for diffs, patches, and repo history. Raw source files do not cover the syntax and distribution agents produce when editing code.

Initial sources:

• SWE-bench issue/patch data: https://github.com/SWE-bench/SWE-bench
• OpenHands SWE-rebench trajectories for final diffs and tool traces.
• BigCode CommitPack metadata: https://huggingface.co/datasets/bigcode/commitpackmeta
• GitHub PR/commit data where licenses and leakage constraints are acceptable.

Break out slices for raw git diff, commit-message-plus-diff, PR-review-plus-diff, and issue-to-patch contexts.

Definition of Done

• Add raw eval builders for at least two diff/patch sources.
• Mask provenance fields that should not be predicted.
• Produce separate metrics for patch text versus surrounding issue/review text.
• Add docs for train/eval leakage checks.

[claude]

🤖 Plan:

Mirrors sibling PR #5084 (which handled the parallel issue #5056). That PR added only the registration scaffolding — a prefix helper and an empty ACTIVE_* registry — and deferred slice population, docs, and pilot gap reports to follow-ups. Applying the same minimal scope here.

• Fetch issue, read skills, preflight duplicate work
• Add experiments/exp5095_diff_patch_ppl.py with DIFF_PATCH_PREFIX = "diff_patch", ACTIVE_DIFF_PATCH_DATASETS, prefixed_diff_patch_validation_sets, diff_patch_raw_validation_sets
• Add tests/evals/test_exp5095_diff_patch_ppl.py (2 tests, both pass locally)
• ./infra/pre-commit.py --all-files --fix — OK
• PR opened: [evals] Add diff/patch PPL registration helper #5103
• Verify CI

Follow-ups (tracked separately, out of scope here): raw-dev builders for SWE-bench / CommitPack / OpenHands diffs, provenance masking, separate patch-vs-surrounding-text metrics, leakage-check docs. Those depend on the license/provenance inventory in #5100.

Relevant code

• experiments/exp5056_raw_web_markup_ppl.py — template from PR [evals] Add raw web/markup/image-text PPL registration helper #5084
• lib/marin/src/marin/evaluation/perplexity_gap.py#L37-L80 — RawTextEvaluationDataset, raw_text_dataset
• experiments/evals/long_tail_ppl_runnable.py — family-grouped slice pattern for later population