[4/4] Stop persisting current_loc (+2% parse throughput) - #919

This change makes it more likely the compiler will enregister the next tape write location (a very hot variable) by making it temporary to stage 2 (i.e. in structural_parser) instead of persisting it in dom_parser_implementation.

This is the last of a series of 4 PRs which simplifies a lot of stage 2 and results in a performance improvement at the end.

Performance

It looks like an overall average throughput improvement close to 2%. The improvements in instruction count and throughput are both pretty much universal. canada.json loses 6% or so overall, but other number parsing things gain substantially. Looking at the numbers, everything has a slightly larger number of cache references, and canada.json is large enough that almost all of them are cache misses, accounting for the difference. The increased number of cache references absolutely does not accord with my theory of why the patch is faster. I'll experiment with that, but I consider this loss acceptable (and recoverable as we delve into streaming in earnest).

PR Notes:

• This change doesn't have its awesome effect on master (it makes it worse). It doesn't get really good until [3/4] Remove unneeded structural_iterator variables #918--presumably because removing several variables from structural_iterator reduced reduced the number of decisions the compiler had to make about registers.
• number_writer: After moving current_loc to the structural_parser, the number_writer couldn't get to it anymore.
  • I couldn't change the dom_parser_implementation &parser to structural_parser &parser, because number_writer gets passed to parse_large_integer and slow_parse_float, and that would deenregister the whole thing.
  • So I changed current_loc from uint32_t to uint64_t * and had number_writer store that. That single pointer is enough to write values to the right place and update the pointer.
  • Unfortunately, that wasn't enough: any attempt to pass next_tape_loc by reference or pointer caused it to de-enregister (reasonably enough). So I pass it by value, let it write the number out, and then bump next_tape_loc after calling either slow_parse_float or parse_large_integer.
• I also changed number_writer to tape_writer and had structural_parser use it everywhere, so the write logic is largely localized.