DocRepoIngester: parallelize LLM extraction across files

[itdove]

Summary

DocRepoIngester.scan_and_process() processes files sequentially in a loop. Each file triggers one or more LLM calls via provider.extract_from_chunk(). Since LLM inference is network-bound I/O, files could be processed in parallel.

Fix

Use concurrent.futures.ThreadPoolExecutor in scan_and_process() to process files concurrently. Cap concurrency to avoid overwhelming the LLM server.

Acceptance Criteria

• scan_and_process processes files concurrently
• Configurable max workers (default: 4)
• Error handling per file preserved (one failure doesn't stop others)
• Results are deterministic (sorted by file path)
• All existing tests pass

Discovered during /simplify review of #37.