C++: Split IRBlock index from IRBlock membership

[jbj]

This is the same change as #123 but for the IR. I still don't know exactly why it's faster, and I invite others to try and replicate my benchmark. I'm using the following query to compare timing and correctness:

import semmle.code.cpp.ir.IR

from int i
where i = any(IRBlock b).getInstructionCount()
select i, strictcount(IRBlock b | i = b.getInstructionCount())

In the QL4E I'm running, the time spent on getInstruction is too low in the "Clause timing report" at the bottom of the log. To see how long the predicate or its replacements actually take, subtract the timestamps before and after in the log output.

[jbj]

Rebased to fix conflicts.

[dave-bartolomeo]

LGTM
If you've measured a speedup for this change on a real snapshot, and you've also measured a speedup for the corresponding non-IR change, I'm willing to believe that the speedup is real. The plausible explanations for why it might not be a real speedup (e.g. slightly reduced memory consumption eliminated swapping) were unlikely to hold for both changes.

[jbj]

Yes, I tested both this and #123 independently.