Add note about JSON tests maybe being wrong.

- Add note about the JSON test maybe not being correct yet.
- Add test to checks the generated names for double underscores to be sure they
  are what is expected.

This should fix the failures in the conformance tests - although
it highlights the problem that we need to do this when changing
the conformance.proto file...

Incase the different runtimes have different generation/implementations for
for the different primitive field types, cover a larger range of the proto
types in the oneof zero tests.

can reasonably be expected to be interoperable.

(The fact that not all Int64/Uint64 values *can* be exactly represented in
IEEE-754 and thus interoperably amongst JSON implementations is precisely the
reason for wrapping them when we serialize... it shouldn't be a surprise that
we can't round-trip in unwrapped form for all values.)

Python 2.x doesn't detect unpaired surrogates so we have to
do that manually.

We should now have no conformance failures.

The tests are run from Travis in the same way as on other
platforms. Currently some expected failures - but only expected
in that they're what we got to start with. Will try to fix them in
other pull requests.