A 32-bit target _could_ create a giant `BigUint` over 2²⁹ bytes = 512MB,
not that this would be computationally useful, but that value would have
more than `usize::MAX` bits. To avoid that possibility of overflow, we
can better represent bit counts in `u64`.

For 64-bit targets, a `BigUint` over 2⁶¹ bytes is not realistic.

This changes the public API in the return values of `bits()` and
`trailing_zeros()`, as well as the `bit_size` parameter of `RandBigInt`
and `RandomBits`. Trying to randomize an absurd size will naturally
cause a capacity overflow, just as if you'd made a huge `Vec`.