perf: use ProcessPoolExecutor for multi-core fuzzy matching

[jakebromberg]

Summary

• Switch Phase 4 fuzzy matching from ThreadPoolExecutor to ProcessPoolExecutor with fork context for true multi-core parallelism (the GIL was serializing threads on a single core)
• Use smaller chunks (~200 artists) for more frequent progress logging
• Add throughput (artists/s) and ETA to Phase 4 chunk logs, elapsed time to Phase 3 log
• Set PYTHONUNBUFFERED=1 in run_step subprocess environment so log lines stream immediately instead of being held in full-buffering pipe buffers

Closes #46

Test plan

• New tests verify ProcessPoolExecutor worker produces same results as direct call
• New tests verify multi-chunk aggregation matches single-batch results
• New test verifies worker initializer sets module globals correctly
• New test verifies Phase 4 logs contain throughput metrics
• New tests verify PYTHONUNBUFFERED=1 is set in subprocess env and caller env vars are preserved
• All 473 existing unit tests pass with no regressions
• Run full pipeline with --prune and observe multi-core CPU usage and frequent chunk progress logs