Allow common process_escapes to handle \x sequences

[KerfuffleV2]

For example, you can do something like \xf0\x9f\x98\x82 to enter 😂 with these changes.

[cebtenzzre]

Since this is C++, why not something like this?

Suggested change

	const char x[3] = { input[input_idx + 1], input[input_idx + 2], 0 };
	char *err_p = nullptr;
	const long val = std::strtol(x, &err_p, 16);
	if (*err_p == 0) {
	std::string x(input + input_idx + 1, 2);
	size_t pos = 0;
	const long val = std::stol(x, &pos, 16);
	if (pos == x.length()) {

[KerfuffleV2]

Well, the first reason is I didn't know to write it that way. :) I think there's a reason to keep it as-is though, creating a const 3 character array probably has lower overhead than making and destroying a std::string, concatenating some other strings, etc. It's probably a small difference, but hey, when I'm feeding a 100,000 token prompt to a 128K YaRN model maybe it'll be .5sec faster.

[cebtenzzre]

Because of small-string optimization there won't be any heap allocation, but I suppose std::string is still slightly more expensive.

If we keep the C version, we should replace const char x[3] with just const char x[] since the size is inferred.

C++ has had std::string_view and std::from_chars since 2017, but we can't use them :/

[KerfuffleV2]

I realized the previous code didn't correctly handle a case where the second hex digit was \0. This version is a little simpler and clearer also. However, I kept the [3] since I feel better about explicitly setting the size when doing pointer arithmetic on the array like that.