The safety guarantees that aren't validated immediately have been
removed (so only keeping the unsafe code in the hi bit extraction, which
is checked right in the call), and this leads to ~3% degradation in
performance in the worst cases, which is still 90% faster than core so
it's well worth it. The unsafe code in the multiplication algorithms is
the cause of the slowdown, but well worth it.

Seems to improve performance on some edge case floats invoking this path
by up to 30%.

Introduce a self-contained `Buffer` trait that allows simpler comparison
of values, and then remove more unsafety, especially in cases where the
compiler can trivially provide the unsafety is sound. Then, also fix a
minor bug with formatting where the mantissa radix negative sign check
was used and not the exponent one.