ufs: do not leave around empty buffers shadowing disk content

If the ffs_write() operation specified to overwrite the whole buffer,
ffs tries to save the read by not validating allocated buffer. Then
uiommove() might fail with EFAULT, in which case pages are left zeroed
and marked valid but not read from the disk. Then vn_io_fault() logic
retries the write after holding the user pages to avoid EFAULTs. In
erronous case of really faulty buffer, or in contrived case of writing
from file to itself, we are left with zeroed buffer instead of valid
content written back to disk.

Handle the situation by releasing non-cached buffer on fault, instead
of clearing it. Note that buffers with alive dependencies cannot be
released, but also either they cannot have valid content on the disk
because dependency on data buffer means that it was not yet written, or
they were reallocated by fragment extension or ffs_reallocbks(), and are
already fully valid.

Reported by: kevans
Discussed with: mav
In collaboration with: pho
Sponsored by: The FreeBSD Foundation
MFC after: 1 week