Retire promises that reference only "new" other promises

[FUDCo]

Paraphrased from the discussion in #1049:

Drop references that have no already exported promises:

This is the "95% solution". The general idea is that resolutions which only reference brand new promises cannot introduce cycles that involve previously-existing promises, and wouldn't cause a recursion problem. Dean said a lot of ERTP/Zoe functions will return a record full of (new) Promises, and this would let their result promises be retired.

For the sending side (vat does syscall.fulfill*), the vat looks at the liveslots tables to see whether the referenced promises are already present or not. It's not immediately obvious how we'll implement this: Dean suggested measuring the size of the tables before+after serialization and seeing if it changed, but we use the same table for both objects and promises. Perhaps we modify the tables to maintain a count of just promise allocation. Or we introduce a generation number and a WeakMap that maps each Promise to an integer, and then after serialization we look at generationMap.get(slotToVal.get(promiseSlot)) > previousGeneration to see if anything is new.

On the receiving side (kernel does dispatch.notifyFulfill*), it's a bit easier, because all the kernel is doing is a linear mapping of slot IDs (kpid and koid) into vat-side references (vpid and void), and it can just ask the c-list if the kernel-id is present before asking it to provide a new one.

[warner]

This is done, we now have the 100% solution (retire all resolved promises, not just the "easy" ones)