r352079 - [FileManager] Revert r347205 to avoid PCH file-descriptor leak.

This is kind of expected: r347205 was a bugfix.
Reverting it causes a real bug, and the tests are failing. The bug is acceptable for 8.0 - it affects only a marginal case (go-to-definition on a #include outside the preamble section of a file that was also transitively #included in the preamble).

One mystery: why are the tests failing on the branch but not on trunk?
I do want to dig into this today/tomorrow, but I don’t think it affects what we do on the branch.

The right fix for the branch is just to delete the three failing assertions (for locations 5, 6, 7 in that test).
Is it possible to do this directly on the branch (without deleting them on trunk)?

Thanks for looking into this. I've reverted the test assertions on the
branch in r352354.