rm: fix diagnostics on I/O error

rm: fix diagnostics on I/O error

I ran into this problem when attempting to recursively
remove a directory in a filesystem on flaky hardware.
Although the underlying readdir syscall failed with errno == EIO,
rm issued no diagnostic about the I/O error.

Without this patch I see this behavior:

  $ rm -fr baddir
  rm: cannot remove 'baddir': Directory not empty
  $ rm -ir baddir
  rm: descend into directory 'baddir'? y
  rm: remove directory 'baddir'? y
  rm: cannot remove 'baddir': Directory not empty

With this patch I see the following behavior, which
lets the user know about the I/O error when rm tries
to read baddir's directory entries:

  $ rm -fr baddir
  rm: cannot remove 'baddir': Input/output error
  $ rm -ir baddir
  rm: cannot remove 'baddir': Input/output error

* src/remove.c (Ternary): Remove.  All uses removed.
(get_dir_status): New static function.
(prompt): Last arg is now directory status, not ternary.
Return RM_USER_ACCEPTED if user explicitly accepted.
All uses changed.
Report any significant error in directory status right away.
(prompt, rm_fts): Use get_dir_status to get directory status lazily.
(excise): Treat any FTS_DNR errno as being more descriptive, not
just EPERM and EACCESS.  For example, EIO is more descriptive.
(rm_fts): Distinguish more clearly between explicit and implied
user OK.
* src/remove.h (RM_USER_ACCEPTED): New constant.
(VALID_STATUS): Treat it as valid.
* src/system.h (is_empty_dir): Remove, replacing with ...
(directory_status): ... this more-general function.
All uses changed.  Avoid undefined behavior of looking at
a non-null readdir pointer after corresponding closedir.
* tests/rm/rm-readdir-fail.sh: Adjust test of internals
to match current behavior.