Don't include A.thy loading time in benchmark

[Chris Warburton]

It can take a while to load the complete A.thy theory, and since it
contains all of our (hundreds of) definitions, we shouldn't include it
in the benchmark times for samples. We *know* that loading hundreds of
definitions can take a while, but that's irrelevant when we're measuring
how long it takes to explore (say) *three* definitions.

Isabelle does time how long it takes to load the A.thy and the
ISACOSY.thy theories separately. I don't think we should rely on the
latter for our timings, but I think it's reasonable to subtract the
former from our own (ASV-based) measurements.

In any case, we should log these as part of the JSON structure, commit
it to git, and do any calculation after the fact.