fix lpad/rpad for zero-textwidth padding

[stevengj]

Hopefully fixes test failure caused by backwards-incompatible change in #39044. Alternative to #41131.

[StefanKarpinski]

The only part I'm unsure about is switching to measuring the string's size with length when the padding has textwidth zero. It is the most backwards compatible, but it feels a bit iffy to me. I guess it does signal that you are not intending to use the function in the "text width mode".

[DilumAluthge]

CI should get fixed on master with #41142, so there's no immediate rush on this.

[stevengj]

One could speculate that anyone who is using lpad or rpad to pad with \0 to a fixed number of "characters" probably actually wanted a fixed number of bytes (as in the Tar.jl case), so their old code was problematic (buggy if there is any possibility of non-ASCII data) and perhaps it is better to throw an error.

On the other hand, more backwards compatibility is good too in 1.x, and from that perspective anyone explicitly using lpad(s, n, '\0') would probably rather get the same answer as in 1.6 than an error ("darn it, I know my s is ASCII") .

[stevengj]

So what is the conclusion here?

[vtjnash]

This is probably surprising and non-reliable behavior, so we probably want to stick with the current behavior (since the Tar bug is now fixed)

[stevengj]

Okay.