Improve calculation of the scale parameter for the uniform float distribution.

[WarrenWeckesser]

The calculation of scale for the uniform float distribution is moved to its own function and updated to avoid the problem that was pointed out in gh-1299. As part of the new implementation, the new utility methods lt_mask(), increase_masked() and utils_next_down() are added to the FloatSIMDUtils traits.

The improved method for computing the scale is explained in the comments in the function compute_scale().

[dhardy]

Quick question from a bird's eye perspective: decrease_masked is used in three locations, but this only replaces one. Is it relevant elsewhere? (Also, is it even needed in new_inclusive?)

[WarrenWeckesser]

In the few tests that I did, new_inclusive didn't show the behavior reported in gh-1299, so I focused on fixing the behavior in new, and didn't look much beyond that. I'll take another look, and also I'll get caught up on gh-1289 and the paper linked there.

In the meantime, I'm marking this PR as draft.

[GUIpsp]

Hello - sorry for the late comment.

I don't have an example off the top of my head, but I can say that I saw the same behaviour with new_inclusive (but with different constants).

[WarrenWeckesser]

Thanks, @GUIpsp. If you happen to find example of the bad behavior with new_inclusive, be sure to note it here or in #1299.

[dhardy]

@WarrenWeckesser is this still on your to-do list?