Apparently I missed this when refactoring a do/while loop into a
standard while loop. It was a harmless extra invocation to check if the
stack was empty.

By default Jiffy is quite strict in what it encodes. By default it will
not allow invalid UTF-8 to be produced. This can cause issues when
attempting to encode JSON that was decoded by other libraries as UTF-8
semantics are not uniformly enforced.

This patch adds an option 'force_utf8' to the encoder. If encoding hits
an error for an invalid string it will forcefully mutate the object to
contain only valid UTF-8 and return the resulting encoded JSON.

For the most part this means it will strip any garbage data from
binaries replacing it replacement codepoint U+FFFD. Although, it will
also try and the common error of encoding surrogate pairs as three-byte
sequences and reencode them into UTF-8 properly.

There was a bug that was resetting the extra amount of space needed when
encoding unicode characters with the `uescape` option enabled.

Numbers like 1.0 were being encoded as <<"1">> which can lead to a bit
of confusion. This merely checks if a decimial point exists and if not
it appends ".0" to the value.

jiffy:encode/2 now accepts a 'pretty' option that will format the JSON
output using two-space indents.

The encoding length for unicode values was applied to the wrong index
variable.

Thanks to Paul Guyot for the detailed bug report and fix.

Fixes #7

I was accidentally truncating numbers when casting to long on
machines that don't have longs that are 8 bytes.

Fixes #5

Based on the buffer doubling scheme in Yajl. This seems to make
Jiffy's encoder times less spikey, but I'm still a bit slower. It
appears to be related to floats or memory handling. Not sure how
to track this down.

The encoder can now return \u escaped unicode data instead of leaving
it as UTF-8 byte sequences. This done like so:

    Eshell V5.8.3  (abort with ^G)
    1> jiffy:encode(<<240, 144, 129, 128>>, [uescape]).
    <<"\"\\uD800\\uDC40\"">>

The encode and decode functions now return the value directly without
being wrapped in a tuple on success. If there is an error, it is
thrown. This is to more closely match the semantics of term_to_binary
and binary_to_term.

Feed the lawyers.

Any number that can't be decoded in C is now passed back
to Erlang for decoding.

Large numbers passed to the encoder will make it through
and be processed in Erlang after the main encoding
process.

I was being an absolute moron and not adjusting my two access
pointers to the data buffer after a realloc.

* Refs became atoms to make sure they can live across calls
  to the NIF functions.
* Initialized curr in decode so that I'm no longer pushing
  random values into the Erlang VM.

Passes all eep0018 tests.