Patch from Dave Miller.  Fixes a very long-standing bug.

If a process has an fd open against a now-removed directory, lookups on that
fd will end up calling ext2_find_entry() against a zero-length directory.

When this happens ext2_find_entry() will, on the first pass through the loop,
set `kaddr' to page_address(page) - 20.  Things get confused and the "zero
length directory entry" warning triggers.

This only happens on 64-bit machines, because ext2_last_byte() is returning
an unsigned (32-bit) value, and the arithmetic works out OK for 32-bit
machines.

So we change ext2_find_entry() to bale out immediately if the directory is
zero-length.  All other directory-walking functions do this, but
ext2_find_entry() forgot to, due to the search-from-the-last-place
optimisation.