Render the history plots in parallel

[sjoelund]

Almost all of the time all-plots.py takes is matplotlib writing svg files: about 0.42s per library for the two plots it draws, near enough independent of how many runs are in the history. Over ~45 branches of ~150 libraries that is more than half an hour of one core, and no library's plot has anything to do with another's.

Hand the rendering to a process pool (-j, all available cpus by default). The queries stay in the main process, and the jobs are generated branch by branch, so the next branch is queried while the previous one is still being drawn; at most 4 jobs per worker are outstanding, as each one carries the whole history of a library.

Everything moves into main() behind an main guard: python 3.14 starts a worker with forkserver, which re-imports the module, and parsing the arguments and connecting to the database again there is not what we want.

Each branch now prints how long its query took, so the log says whether it is the queries or the rendering that is slow.

Measured on a synthetic 80-library database:

jobs  time
   1  28.2s
  12   7.7s
  24   4.0s

The svg files are identical apart from the embedded date and matplotlib's per-run element ids, which differ between two serial runs as well.

Assisted-by: Claude Opus 5 (1M context)