Namely, we don't require that a number be exactly representable as a
single precision float anymore (this requires many `Math.fround` calls).
There is no precision loss risk (compared to the `long` use-case, which
requires explicit bounds), since we actually get bonus precision (via
JavaScript's use of doubles); we simply transparently return back to
floats when serializing.

Also change how boolean validity checks are performed (equivalent but
likely a bit more efficient).