Fix sample-limited dataset extractors

[neubig]

Summary

• remove hardcoded AgentTuning sample ID filters so extract_raw emits all valid examples for AgentInstruct configs
• remove the Android-in-the-Wild five-record per split cap
• make AndroidControl ignore the bundled tf_record_sample and use local full TFRecords or the public GCS full-corpus glob

Validation

• python -m py_compile on all touched extract_raw.py files
• static grep confirmed the AgentTuning SAMPLE_IDS gates and Android-in-the-Wild i >= 5 cap are gone
• runtime AgentTuning smoke test was attempted, but local Hugging Face datasets loading failed in the existing venv while reading cached feature metadata before extractor iteration

Context

The ADP v2 upload included several configs with only 3-5 rows because these extractors were sample/demo-limited while the upload pipeline treated completed artifacts as full datasets.

Tracking

Related to #218.

Live evidence

• 2026-07-14: at PR head 90dbcff, ran the AgentInstruct AlfWorld extractor against the public Hugging Face parquet source. The source generated 336 rows, and the extractor emitted consecutive valid records alfworld_0 through alfworld_5. This exercises the new full-corpus path and demonstrates that the previous five hard-coded IDs are no longer the output limit.
• Command: uv run --with datasets python datasets/agenttuning_alfworld/extract_raw.py piped to head for six emitted records; the extractor handled the closed pipe cleanly.

Closes #310