This patch changes how nsConverterInputStream handles passing data
through to the underlying unicode converter in order to make it more
reliably handle propagating errors and deal with short reads from the
underlying input stream.

This was done by making the code continuously read within the Fill
method until at least one character has been decoded from the input
stream, so that we don't spuriously communicate an EOF to the caller due
to a short read not producing enough bytes for the decoder to produce a
UTF-16 character.

In addition, while making this change it became easier to signal to
the decoder about the final read from the input stream, meaning that
partial characters at the end of the stream will now generate a
replacement character, rather than being ignored.

Differential Revision: https://phabricator.services.mozilla.com/D152682