Explore replacing Nat::from_float with a ctor that takes double

Need to regression test this along with uintmax_t ctor to make sure neither tread on eachother's toes.

Currently, trying to construct Nat from double directly will bind to uintmax_t ctor. That's fine, but we don't want to inverse to happen (casting uintmax_t to double will lose precision!) so we should test both. In an ideal world, we'd have both as separate ctors.