Range cannot be inclusive for floats

[vks]

Range::new and Range::new_inclusive are currently identical for f32 and f64.

[vks]

It is not clear whether this can be corrected (see dhardy#85 (comment)), but we should at least document it or even panic on RangeFloat::new_inclusive.

[pitdicker]

We already have the following documentation:

Types should attempt to sample in [low, high) for Range::new(low, high), i.e., excluding high, but this may be very difficult. All the primitive integer types satisfy this property, and the float types normally satisfy it, but rounding may mean high can occur.

So we already say sampling from floating point ranges might be fuzzy at the boundaries (as users probably expect as rounding errors are part of the deal with fp). Do you think we really need to spell things out even more? We could turn the comment from that method into a doc comment...

We should definitely not panic. There is no programmer error here and no incorrect results. And we don't want to make using RangeFloat::new_inclusive hard to use generically.

Of course, trying to reduce the rounding error to be more accurate at the boundaries can always be good, but I suspect there is no silver bullet that does not decrease the accuracy in other places.

[dhardy]

Indeed, I don't think there's much we can do here.

@vks if you want to make a PR improving doc please do, or we can also reopen this, but I don't see anything constructive here.