benchmark

tools/compare: don't actually discard valid (but zero) `pvalue` (#1733)

* tools/compare: when dumping json, pretty-print it

It's rather completely non-human-readable otherwise.
I can't imagine the filesize really matters,
and if it does, it should just be compressed later on.

* tools/compare: add failing test

* tools/compare: don't actually discard valid (but zero) `pvalue`

So, this is embarressing. For a very large number of repetitions,
we can end up with pvalue of a true zero, and it obviously compares false,
and we treat it as-if we failed to compute it...