Simplify Dict representation

[oscardssmith]

remove idxfloor and maxprobe, and shrink age to UInt32.

idxfloor is removed because the only thing it helps with is speeding up iterate calls (very slightly since it only speeds up the first call), and the speedup comes at a cost on insertions and deletions that happen to land near the front of the dict.

maxprobe is removed because if the hash function is good, we know that the Dict will spread values well, and if the hash function isn't good, rehashing won't fix anything, so we spend a lot of effort tracking maxprobe, but in most circumstances, it won't do anything, and in the circumstances where it does, it probably would be better off if it didn't.

Rebased and slightly more aggressive version of #48380 (rebasing was going to be more annoying since that PR was pre Memory).

[giordano]

Are these effects needed now that you have a fixed iteration instead of while true?

[oscardssmith]

unfortunately, yes. the compiler doesn't know that fixed length iteration terminates

[KristofferC]

Rebased and slightly more aggressive version of #48380 (rebasing was going to be more annoying since that PR was pre Memory).

FWIW, you can also just force push (instead of rebasing) to the existing PR to not lose the relevant discussion on it.

[oscardssmith]

I talked to @vtjnash today, and he pointed out that the point of idxfloor is to prevent quadratic behavior when deleting all entries from a Dict via repeated pop! calls.

Given this, this PR as is isn't good as is. If we wanted to remove idxfloor, we would also have to sometimes shrink the dictionary on deletions if the count was too small (to e.g. ensure something like size(table)/8 < count < 4/5 *size(table)).

As such, this will some substantial benchmarking to determine whether the right approach is to remove idxfloor and resize on deletion, or whether idxfloor needs to stay.