Automate shrinking benchmark more

[tybug]

Brings more of the shrinking benchmark into pytest for automation. Also replaces our plotting code with a vega specification. The plotting code requires a dependency on vl-convert to generate the plot. I know about altair, but to my understanding that only aims to represent vega-lite specs, and we are moderately abusing vega (not lite) to have two axes.

Here's a benchmark result where old == new, which gives an idea of the variability in current shrinking. 10 trials (5 old, 5 new), 95% CI. blue is absolute difference, red is relative difference. (@DRMacIver you might find this interesting!)

[tybug]

I'm happy to report that putting try_trivial_examples first in #4203 was the right decision! Though I would warn against putting too much stock in this benchmark, as our tests are likely biased more towards unusual edge cases than normal shrinks.

try_trivial_examples first	try_trivial_examples after node_program("X")

Graphs are relative to the shrinker without try_trivial_examples.