While working on OCP I waited four hours for every CI run. Generating OCP with pywrap was the first hour of it, three times over, one per platform. Generating OCP (320 modules, ~7000 headers) on a 4-core runner went from 49 minutes to 12 on Linux, and about the same for Windows. The OSX bindings took 169 minutes on a mac runner and take 15 once they are generated on Linux as well (OCP side in CadQuery/OCP#229). Four commits:

• PCH of the shared preamble, parsed once per worker: parse 32 -> 10 min per platform, output unchanged.
• Symbol lookup by bisection instead of a table scan per method: transform 27 min -> 28 s on OSX.
• Collections chunks of 25 instead of 100: the largest translation unit 5 -> 3.2 GB, which is what lets OCP compile with four jobs.
• --target=x86_64-apple-darwin and the platform settings taken from the target's section, so the OSX bindings generate on Linux.

Found on the way and fixed:

• the platform settings (prefix, includes, parsing_header) came from the host section, not the target's.
• the default args list of parse_tu grew with every call.