The word-diff logic accumulates + and - lines until another line type
appears (normally [ @\]), at which point it generates the word diff.
This is usually correct, but it breaks when the preimage does not have
a newline at EOF:

  $ printf "%s" "a a a" >a
  $ printf "%s\n" "a ab a" >b
  $ git diff --no-index --word-diff a b
  diff --git 1/a 2/b
  index 9f68e94..6a7c02f 100644
  --- 1/a
  +++ 2/b
  @@ -1 +1 @@
  [-a a a-]
   No newline at end of file
  {+a ab a+}

Because of the order of the lines in a unified diff

  @@ -1 +1 @@
  -a a a
  \ No newline at end of file
  +a ab a

the '\' line flushed the buffers, and the - and + lines were never
matched with each other.

A proper fix would defer such markers until the end of the hunk.
However, word-diff is inherently whitespace-ignoring, so as a cheap
fix simply ignore the marker (and hide it from the output).

We use a prefix match for '\ ' to parallel the logic in
apply.c:parse_fragment().  We currently do not localize this string
(just accept other variants of it in git-apply), but this should be
future-proof.

Noticed-by: Ivan Shirokoff <shirokoff@yandex-team.ru>
Signed-off-by: Thomas Rast <trast@student.ethz.ch>
Signed-off-by: Junio C Hamano <gitster@pobox.com>