Probably missed as part of some recent header cleanup.
Inspired by the discussion on D118694, this adds a straightforward fuzzer for bitcode reading. Currently it will very quickly run into OOM, because we do unconditional vector reservations with user-provided sizes.