Improve code completion in failure cases in two ways:

  1) Suppress diagnostics as soon as we form the code-completion
  token, so we don't get any error/warning spew from the early
  end-of-file.
  2) If we consume a code-completion token when we weren't expecting
  one, go into a code-completion recovery path that produces the best
  results it can based on the context that the parser is in.

llvm-svn: 104585