Put inner multireweight script in CONVERGECAST-COLLECT-ROOTS

[karalekas]

Also:

• Make sure to use loose address equality in blossom-let
• Fix a bug in one of the MRW tests
• Need bleeding edge aether now for tests to pass

[ecpeterson]

I don't really understand the semantic difference here. Is there one? The only change of that sort I see is the missing aborting? check that used to wrap the "inner" routine.

Looking at this, I feel that "inner" was a misnomer — we haven't yet acquired the locks, so how could we be in the critical section? — so I don't mind renaming / reorganizing it. If all the other operations do have critical sections demarcated as inner subroutines (which I think is true? at least all the primary ones do, though maybe not reweight...), then I'd prefer that this one still have a (properly defined) inner subroutine too. But none at all is better than an improper one.