Basically, the algorithm to append a converted string to an existing
`smart_str` works by increasing the `smart_str` buffer, to let `iconv`
convert characters until there is no more space, to set the new length
of the `smart_str` and to repeat until there is no more input.

Formerly, the new length calculation has been wrong, though, since we
would have to take the old `out_len` into account (`buf_growth -
old_out_len - out_len`).  However, since there is no need to take the
old `out_len` into account when increasing the `smart_str` buffer, we
can simplify the fix, avoiding an additional variable.

We must not ignore erroneous characters in mime headers, but rather let
iconv_mime_decode() fail in this case, issuing the usual notice
regarding illegal characters.

We have to cater to the possibility that `=?` is not the start of an
encoded-word, but rather a literal `=?`.  If a line break is found
while we're still looking for the charset, we can safely assume that
it's a literal `=?`, and act accordingly.

If we're expecting the start of an encoded word (`=?`), but instead of
the question mark get a line break (CR or LF), we must not append it to
the `pretval`.

Fix duplication of recursively checked deps

There's no need to actually try to trigger an out-of-memory condition
to proof the leak; instead we can simply rely on the Zend MM to report
the memory leaks in debug mode (at least on Linux).  Therefore we
simplify the regression test, which also makes it run much faster.