Add opt-in deduplication to the MedDialog scenario

[azrabano23]

Problem

The upstream med_dialog data contains duplicate entries; the icliniq test split has 3108 entries but only 2069 unique source texts (#3746, confirmed by @yifanmai).

Approach

This implements the deduplication inside HELM, following the review feedback on the previous attempt (#4178), which stalled after changes were requested:

• MedDialogScenario gains a dedup argument that defaults to False, and get_med_dialog_spec() exposes the same argument — exactly as requested in that review, so previously published runs remain reproducible bit-for-bit. With dedup=False (the default) the emitted ScenarioSpec and RunSpec are identical to today's.
• When dedup=True, examples whose src duplicates an earlier example are dropped, keeping the first occurrence in the original order, and the run spec name becomes med_dialog,subset={subset},dedup=True so deduplicated runs are distinguishable from existing ones — the "new version of this scenario" suggested in the issue thread.

Tests

New test_med_dialog_scenario.py with two tests:

• test_med_dialog_scenario_dedup — deterministic, no network (it pre-seeds the split file, so ensure_file_downloaded skips the download; runs in default CI): asserts the default keeps every entry including duplicates, and dedup=True keeps exactly the first occurrence of each source text in the original order, with the matching references.
• test_med_dialog_scenario_dedup_real_data — marked scenarios: asserts the real icliniq test split yields 3108 instances (2069 unique) by default and 2069 unique instances with dedup=True, preserving first occurrences.

pytest src/helm/benchmark/scenarios/test_med_dialog_scenario.py                # deterministic test
pytest -m scenarios src/helm/benchmark/scenarios/test_med_dialog_scenario.py  # real-data test

Note: at the time of filing, the CodaLab Worksheets REST API is returning HTTP 500 for this bundle (and for bundle queries generally), so the real-data test — and any fresh run of this scenario on main — currently cannot download the data. The counts in the real-data test are the ones reported and confirmed in #3746.

Closes #3746.