We can enhance our fuzz-testing coverage by hooking Encoder also into
the loop: if input data is suceessfully decoded, we have an object that
can be passed back to Encoder to generate a pickle. We can't tell that
that pickle must be the same as original input data, since pickle
machine allows multiple representations of the same data. However we can
assert that when that pickle is decoded back it should be the same as
encoded object.

This catches several problems:

- marker is currently returned as pickle result (see next patch).
- text-based STRING and UNICODE are not properly decoded (no fix yet).
- self-referencing data structures kill Encoder (no fix yet).