Fixed on the fly decoding of base64 attachment. We need to do on the fly

Fixed on the fly decoding of base64 attachment. We need to do on the fly
decoding because otherwise we have to buffer entire attachments in memory
which will cause out of memory errors with large attachments.

The fix has to do with the grouping of bytes in pairs of 4 which represents
24 bits of data. If such group is splitted over 2 lines then on the fly
decoding went wrong. Now we count the pure base64 data (without \r\n, \r \n
and spaces), divide it by 4 and return incomplete pairs which get processed
the in the next loop.

git-svn-id: https://svn.code.sf.net/p/squirrelmail/code/trunk/squirrelmail@7343 7612ce4b-ef26-0410-bec9-ea0150e637f0