Tweak test harness to handle failures correctly

Add support for Python 2.6 and 3.3

add sort_keys

Feature/write null values as comments

Strip underscores for int() and float()

Fixes a parse failure on integers/floats with underscores.

While the script has no problem detecting numbers with
underscores, `_p_value` would choke right before returning
them due to the strings being passed unmodified into
`int()` and `float()`.

There is not more lexing involved, all token extraction is now
context-sensitive. This fixes #3.