[Training] Fix collect-and-compile max_workers mismatch

[agorevski]

Evidence

• train.py:133-137 calls builder.collect_and_compile_contracts(..., max_workers=3, max_compiler_configs=...).
• src/dataset_pipeline.py:793-797 defines collect_and_compile_contracts(self, contract_addresses, max_compiler_configs=2) with no max_workers parameter.
• src/dataset_pipeline.py:816-918 then processes contracts, compiler configs, and compiled contracts in nested serial loops.

Impact

The full training pipeline can fail before training starts with TypeError: collect_and_compile_contracts() got an unexpected keyword argument 'max_workers'. Even after the signature mismatch is fixed, dataset compilation/TAC preprocessing remains serial, increasing wall-clock time and compute cost before every fresh training run.

Recommended fix

Accept and honor a bounded max_workers argument in collect_and_compile_contracts, or remove the caller argument if serial execution is intentional. If parallelizing, use per-thread SQLite connections, rate-limit Etherscan requests, and guard solc installation/cache access.

Acceptance criteria

• A unit/integration test covers the train.py collection call path and proves max_workers is accepted.
• Dataset collection can be configured with 1 worker and >1 workers without SQLite/solc races.
• Logs include contract/config throughput so preprocessing speedups are measurable.

[agorevski]

Resolved in commit e4a1530. Validation: python -m compileall -q train.py download_hf_contracts.py src web scripts tests; uv run pytest -q (521 passed).