libext2fs: don't old the CACHE_MTX while doing I/O

libext2fs: don't old the CACHE_MTX while doing I/O

A report a deadlock problem caused by I/O errors (caused by e2fsck's
error handler trying to write to a bad block to perform a forced
rewrite) uncovered that we were holding the CACHE_MTX while doing read
operations.  This serialized read operations which destroyed the
performance benefits from doing parallel bitmap loading (or the
parallel e2fsck processing under development).

So restructure the code in unix_read_blk64() so that the read is
always done into the user-provided buffer, and then copied into the
cache afterwards.

Signed-off-by: Theodore Ts'o <tytso@mit.edu>