Parse streams incrementally.

Feature/small float serialization

Moved buffer management into the lexer class and implemented YYFILL so that streams are read incrementally.

Now uses std::snprintf() to generate a "%.15g" formatted string
for JSON values of type number_float. 15 decimals digits are
enough to round-trip an IEEE 754 double from string->double->string
and get an identical result.

std::snprintf is called twice. Once to determine the required
buffer size and then again after allocating a buffer of that
size.

Note that the buffer size *could* be hardcoded for better
performance. "%.15g" should result in strings of maximum length
23, plus one character for the terminating null for a buffer size
of 24.

template version with re2c scanner