Avoid streaming incomplete UTF-8 characters

[corebonts]

Some characters, like the chinese fù is sometimes returned as two tokens, as "\u00e8\u00b5" and "\u008b" in this case.

This is also depends on the model, but when it happens, for example with DeepSeek R1, we have to wait for the character to be complete and send it only then.

This resolves #722 and #646

[corebonts]

The UTF check is from here:

llamafile/llama.cpp/server/server.cpp

Lines 1032 to 1059 in 33f561d

	// check if there is incomplete UTF-8 character at the end
	bool incomplete = false;
	for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i)
	{
	unsigned char c = slot.generated_text[slot.generated_text.size() - i];
	if ((c & 0xC0) == 0x80)
	{
	// continuation byte: 10xxxxxx
	continue;
	}
	if ((c & 0xE0) == 0xC0)
	{
	// 2-byte character: 110xxxxx ...
	incomplete = i < 2;
	}
	else if ((c & 0xF0) == 0xE0)
	{
	// 3-byte character: 1110xxxx ...
	incomplete = i < 3;
	}
	else if ((c & 0xF8) == 0xF0)
	{
	// 4-byte character: 11110xxx ...
	incomplete = i < 4;
	}
	// else 1-byte character or invalid byte
	break;
	}

[cjpais]

I've tested this and it works. We may want to modify the function slightly later, but for now it works from my basic testing, so pulling it in.

Thanks so much for the contribution!