- Removed copy of field names in binary deserialization
- Removed copy of array values in binary deserialization
- Removed copy of string values in json deserialization
- Removed unhelpful allocation in json string value parsing
- Removed copy of blob data on binary and json serialization
Number matching semantics are slightly changed: since this is used
as a filter to check whether a number is signed and/or floating
point, we can speed this up further. strto* functions are called
afterwards and will error out where necessary. We now also accept
numbers like .4 which were not accepted before.
The strto* calls on a boost::string_ref will not access unallocated
memory since the parsers always stop at the first bad character,
and the original string is zero terminated.
in arbitrary time measurement units for some arbitrary test case:
match_number2: 235 -> 70
match_word2: 330 -> 108
