fix(framework): preserve remote eval row origins

[Curtis Galione (cjgalione)]

Summary

• preserve valid inline row origins through dictionary-to-EvalCase conversion
• prefer a complete active dataset-row identity over preserved source provenance
• propagate the resolved origin to root spans, streamed progress events, and EvalResult
• omit malformed origins and update the relevant public/internal types

Why

Remote playground requests include each selected row's origin, but Python's EvalCase.from_dict() dropped it because EvalCase did not declare the field. Streamed progress events consequently emitted a null origin, preventing the playground from associating results and scores with source rows and making runs appear stuck at 0%.

Validation

• .venv/bin/pytest src/braintrust/test_framework.py -k 'origin' -q
• .venv/bin/pytest src/braintrust/devserver/test_dataset.py -k 'returns_inline_data_unchanged' -q
• .venv/bin/nox -s test_core
• .venv/bin/nox -s test_types
• .venv/bin/nox -s pylint
• pre-commit hooks for the changed files

Tracking

resolves SDK-204