gives a chance to move into a spawn_blocking and perhaps limit big-car concurrency

so we can use all jacq's good firehose stuff but repo-stream for the fun bits

its default hash function isn't DOS-resistant, and we need to be robust by default

doing mst slices should require a prefix

doing sparse msts should be its own thing

we're back (at least with mimalloc)

and then get dragged by the benches

...car slice handling not yet actually implemented

yay we're back to the old api (doctests pass again!!)

try to put the data into a more walk-friendly shape

oops it made things slower

benches:
huge: +0.36% (within noise margin)
midsize: +3.3%
little: +3.7%
tiny: +4.6%
empty: +6.4%

and hashbrowns

looks like another 7% improvement for ~free (allocator joke)

pretty good speedup!