dd: fix nocache regions passed to posix_fadvise()

dd: fix nocache regions passed to posix_fadvise()

Previously with oflag=direct the call to invalidate_cache()
was not passed to the kernel, as it was less than a page size,
and a subsequent call was not made to invalidate the pending space.
Similarly with oflag=nocache the pending space at EOF was
not invalidated.  Even though these amount to only a single page
in the page cache it can be significant.  For example on
XFS before kernel patch v4.9-rc1-4-g0ee7a3f, O_DIRECT files
would have been read inefficiently if any pages were cached,
even if they were already synced to storage.

* src/dd.c (i_nocache_eof, o_nocache_eof): New bools used
to control when we want invalidate_cache(,0) to clear to EOF.
(cache_round): Use IO_BUFSIZE (currently 132KiB) to minimize
calls to the relatively expensive advise function, rather
than page_size.  This also makes it clear that while the
kernel function operates on pages, this size is chosen for
performance reasons.
(invalidate_cache): Refactor to share more code between
input and output paths. Use i_nocache_eof and o_nocache_eof
rather than proxying off max_records.  Ensure we
invalidate full pages when clearing to EOF as the kernel
will ignore any non complete pages.  Fix the offset used
for the output path.
(dd_copy): Invalidate the cache of the input after the
offset is updated, for consistency and so we don't try to
invalidate before the start of the file.  When we read
EOF on input, set flags so that we invalidate to EOF.
(main): Invalidate to EOF in more cases, by depending
on the i_nocache_eof and o_nocache_eof flags.
* doc/coreutils.texi (dd invocation): Clarify the alignment
and persisted caveats on the example applying "nocache"
to part of a file.
* tests/dd/nocache_eof.sh: A new test.
* tests/local.mk: Reference the new test.
* NEWS: Mention the bug fix.
Issue reported by Eric Bergen.