Build daemon for katex

[SquidDev]

I've been spending some time looking at what the slow bits are of a clean 1lab build:

Name	Count	Total	Average	Max
node	1,931	6m18s	0.20s	0.53s
pdflatex	258	4m16s	0.99s	1.80s
agda	1	4m04s	4m04s	4m04s
agda html	227	48.81s	0.22s	1.12s
git	383	12.58s	0.03s	0.13s
pdftocairo	258	9.54s	0.04s	0.22s

While there's some things which are always going to be slow (_^Agda), the obvious issue here is KaTeX. During a fresh build, we're spinning up 1.9k KaTeX processes, each of which runs for a fraction of a second!

It would be worth experimenting with creating some sort of build daemon here; just a basic JS program which reads an equation from stdin and writes the resulting SVG to stdout. Our Shake code can then maintain a pool of these, avoiding the cost of spawning a process for each job.

[SquidDev]

One other thing @plt-amy suggested is kicking off the KaTex and diagram builds earlier. We can create a separate job which parses Markdown and compiles the equations/diagrams, and start that at the same time as the main Agda build.

It won't reduce single-core time, but means we can benefit from parallelism earlier on, meaning we don't end up with a build graph like this: