docs: Modify specification of trunc_u and trunc_s.

[cqc-alec]

Turns out the implementation in the constant-folding code is already in line with this modified specification, so this is not a breaking change.

Closes #771 .

[ss2165]

thanks!

[cqc-alec]

Hmm I may have been a bit hasty -- it looks like rust (at least by default, since 1.45) performs "float-type as integer-type" conversions according to a bunch of rules that include sending non-finite and NaN values to well-defined values, so the implementation is not quite in line with the spec. rust-lang/rust#71269
In two minds as to which way to go.

[cqc-alec]

There's a good summary on that PR of what other languages do, which is quite variable, but I tend to agree with the author's comment that "overall rounding to zero and saturation is the common case", so am inclined to modify the spec accordingly. (This would mean these functions could not produce an error.)

[cqc-alec]

Was misreading the code. There is a check for finiteness. The only edge case remaining is casting a negative float to an unsigned; this will result in 0, which I think is adequately described by "rounding towards zero".