Use target state to compute committees

[paulhauner]

Issue Addressed

Related to #4234, but does not fully resolve it.

Proposed Changes

When we have a shuffling cache miss, sometimes compute the shuffling by loading the target state rather than the beacon block root state.

This is cheaper, because we don't need to apply blocks/skip slots to generate a mid-epoch state. It also avoids weirdness with the incomplete state advances.

Additional Info

NA

[paulhauner]

After discussions with @michaelsproul we found this isn't going to work when shuffling_epoch + 1 > epoch(beacon_block_root) :(

[paulhauner]

I've updated this PR to only sometimes use the target state, since we sometimes need the head block state.

Now this PR doesn't full resolve #4234, but it provides a useful optimization IMO.

[paulhauner]

I'm bumping this to v4.2.1 since it's not so important and v4.2.0 already has a decent amount of complexity.

[michaelsproul]

I'm concerned we have a race condition here that could make this error reachable, resulting in an ugly error! here. The race is:

1. Verify attestation block isn't finalized (it's not, yet)
2. Check fork choice, attestation block is present
3. Finalization updates, and so does the split point
4. The attestation slot is now prior to the split and triggers this error

Alternatively:

1. Verify attestation block isn't finalized (it's not, yet)
2. Check fork choice, attestation block is present
3. Check attestation slot against split slot, still fine
4. Finalization migration runs and advances the split
5. Try to run get_state and fail, because the state has been pruned. Or (today) load a freezer state at great expense ( 😱 ).

This second race is related to #4610.

My preference would be to soak #4160 in Hydra and ensure it's solid before making any changes here. Then we can add this change in maybe with the following defensive changes:

• No new RequestedColdShuffling error which complicates error handling (reuse AttestationStateIsFinalized which already has its own handling).
• No separate split slot check: use get_inconsistent_state_for_attestation_verification_only in both branches and let it do the check atomically (this change exists on the [Merged by Bors] - Remove the unused ExecutionOptimisticForkVersionedResponse type #4160 branch).

[paulhauner]

Good points @michaelsproul. I've dropped the v4.3.1 tag since I don't think this change is urgent. I'll come back to it later, it might even be more trouble than it's worth.