fix overflow in prevfloat with unsigned stepcount

[adienes]

fixes #59661

[adienes]

I think the docs could stand an update, as applying a function -n times doesn't make much sense (although ofc most people will likely understand what it's intending to say)

instead of

nextfloat(x::AbstractFloat, n::Integer)
The result of n iterative applications of nextfloat to x if n >= 0, or -n applications of prevfloat if n < 0.

it should probably be something more like

The result of n iterative applications of nextfloat to x if n >= 0, or if n < 0, return a value p such that prevfloat(p, n) == x

but I don't want to change the docs in this PR as that would likely invite stricter scrutiny

[nsajko]

This change could break prevfloat for a package which implements nextfloat and expects prevfloat to have a generic fall back implemented using nextfloat. So I think this may be a "minor change"?

[nsajko]

I guess you may want to use the new ispositive and isnegative, instead of <.

[adienes]

This change could break prevfloat for a package which implements nextfloat and expects prevfloat to have a generic fall back implemented using nextfloat

those packages were already broken as the existing fallback is incorrect

[vtjnash]

LGTM