Moved all dump()-related functions into a class "serializer". This fix includes a lot of performance improvements yielding a 7% speedup for serialization. Details on the individual steps can be found in the commit messages.

Individual benchmark numbers:

before:

dump jeopardy.json                            5   374555228 ns/op
dump jeopardy.json with indent                5   430953700 ns/op
dump numbers/floats.json                      5   622938509 ns/op
dump numbers/signed_ints.json                20    82177979 ns/op

after:

dump jeopardy.json                            5   335449757 ns/op -11%
dump jeopardy.json with indent                5   375467773 ns/op -13%
dump numbers/floats.json                      5   584611852 ns/op -7%
dump numbers/signed_ints.json                20    68289574 ns/op -17%

- Added comments for the serializer class.
- Added test case for resizing of the indentation string.
- Using std::none_of to check if “.0” needs to be added to
floating-point number.

snprintf returns an int, but we later assign it a difference_type which
is usually a long.

A lot of small changes to avoid memory allocations:

- The locale is only queried once rather than with every number
serialization.
- The indentation string is recycled between different calls.
- The string escape function avoids a copy if no escaping is necessary.
- The string escape and the space function use a complete switch case
instead of cascaded ifs.

Cachegrind measures some 15% performance improvement.

Treated the size of the range as the number of thousand separators.
This logical error yielded a negative value for written_bytes and
eventually an infinite loop, as written_bytes was converted to an
unsigned value.

By merging numtostr into serializer, we can write directly to the
output stream. As a consequence, all stream calls are now unformatted.

The class is currently just a wrapper for an std::ostream and collects
all functions related to serialization. The next step should be
recycling of variables to avoid repetitive initialization for each
recursive dump call.

This reverts commit 909b439b.
For some strange reason, the test suite crashes when compiled
with GCC.

numtostr now directly writes to a stream. Return value of snprintf is
reused to avoid finding end of string. Cachegrind suggests a 1%
performance increase.

All ‘<<‘ calls have been replaced by write()/put() calls. The
indentation strings needs not to be resized. Cachegrind measures 1%
performance improvement.

Indentation string is recycled to avoid allocations. Comma-separation
in objects does not need an if any more. Cachegrind measures 1%
performance improvement.

Added separate code paths for normal output and pritty-printed output.
This allowed to remove most of the ifs along the way. Benchmarks and
cachegrind suggest a 10% performance improvement.

🔀 merge branch 'cotire' of https://github.com/tusharpm/json into tusharpm-cotire

- get CMake for XCode 8.1 image

- Add prefix header
  - Include catch.hpp
  - Include json.hpp
    - Replace private with public for all json_unit files
- Move `unit.cpp` to an object library
- cotire issue: strip whitespace from CMAKE_INCLUDE_SYSTEM_FLAG_CXX