This reapplies r209910 with a fix for the assertion failures hit on the
buildbots.

original commit message:
I thought we could get away without this, but it means that the
FileEntry objects actually refer to the wrong files, since pcms are not
updated inplace, they are atomically renamed into place after compiling
a module.

So we are close to the original behaviour of invalidating the cache for
all modules being removed, but now we should only invalidate the ones
that depend on whichever module failed to load.

Unfortunately I haven't come up with a new test that didn't require
a race between parallel invocations of clang.

<rdar://problem/17038180>

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@209922 91177308-0d34-0410-b5e6-96231b3b80d8