perf: improve build performance for large packages

[rich-iannone]

This PR has several Improvements for build performance (for large packages). There are two optimizations that significantly reduce build times for packages with large APIs:

• export discovery: use cached griffe tree instead of per-item dynamic imports
• parallelized Markdown page generation

Measured against Pointblank (100 exports, 37 classes, 574 source-linked items, 301 pages):

• source links step: ~1m 16s → <1s
• post-render markdown: 75.5s → 12.6s
• importantly, all 10,012 tests pass

Partially addresses: #261