The rodata version is faster, but this version is close enough that it
warrants not having the large amount of rodata included in the overhead.

This caches the previously generated string and tries to handle the common
case of sequential printf() by incrementing the last character. If that
overflows, we walk towards the start of the string until we do not need
to overflow.

This is still roughly 20x faster than using sprintf() directly, as we did
previously, for sequential strings.