Content-addressed bitcode and object files

[woodruffw]

This is still WIP; it probably needs more work + cleanup to bring it to parity with the current behavior.

See #58.

[ianamason]

OK let me know when it's ready. We should make sure the (right version of the) linux kernel still builds.

[woodruffw]

Leaving some notes to myself:

• This approach needs to be a little more thoughtful. It currently it blindly creates a list of ($hash.bc, $hash.o) pairs that never get merged back in, since the ultimate executable is linked against foo.o instead of its $hash.o dual.
• There are probably still some sources of unsoundness left, although they're maybe no more unsound than the misbehaving build system that would trigger them. In particular, a build system that recompiles foo.c concurrently might clobber the content-addressed paths, but such a build would already be at risk of writing clobbered object files.