Add an fmt::print()-like function

[eyalroz]

The fmt library's fmt::print() function targets the stdout stream, so we can't use it. However, we should be able to adapt it, or most of it, to use CUDA's printf() function. Let's do that.

... this is not so trivial, since while fmt may be willing to call system library functions several times, even at the level of a single characters, we mustn't do that - to avoid different threads' printing getting jumbled on the same line. That means we have to be more pro-active with buffering. But that means we're in the dilemma of static vs dynamic sizing of a buffer etc.