Try to diagnose untranslatable input when using iconv

Try to diagnose untranslatable input when using iconv

See Debian bug #348909.

The problem starts with the fact that iconv returns EILSEQ (invalid input)
when in fact the input is merely untranslatable.

It is possible to diagnose this situation by running another conversion with
the output encoding the same as the input (so that it will always succeed on
valid input) at the same point. This is what we now do. Unfortunately,
there’s no way I can see to work out how much input to skip (i.e. the length
of the untranslatable character in the source encoding). Hence, we still
just skip one byte. The typical result is that invalid input is diagnosed on
the next step, resulting in the same problem as at present.

Two possible workarounds are to not use iconv, or to set abort_level to
RECODE_UNTRANSLATABLE (this is what test_2 in t80_error.py does).