benches: Use criterion
Me failing to run enough benchmarks in !5 (merged) lead me to think that perhaps with criterion the lack of change in perf might be easier to spot. Testing !5 (merged) on these commits with similar background noise does yield the No change in performance detected..
Obvious cons with criterion is that it does add even more compilation time to this previously fast to compile crate.
Pros include no longer requiring nightly usage.
Instead of a single long bench for each {buffer,stream} there are now a number of smaller sizes (8-16384 bytes) which are benchmarked. The choice of the buckets was quite arbitary.