Further optimise test runtime, perhaps uselessly

There's maybe no great reason to do this, but this pushes test runtime down a little further still.

Most of the work goes into adding data that we can share between every fixture of the same test run, so we can ameliorate things like loading and parsing PNGs. This comes in handy for output-transform and buffer-transform tests, where we can just have a couple of image loads and SHM allocations, rather than like 56.