Sync committee duty calculation for an epoch is too long (4 seconds)

[Arkenan]

This is a tl;dr on Rodrigo's notes on PR #1297.

On the one hand, this may not be an issue on mainnet, as it's mainly originating from all the validators being part everywhere. This is the core problem:

for 60 validators (the test I was seeing the 4 seconds delay) we have to calculate proofs for every subcommitte (4) for every slot (32). The proof calculation cost approximately 0.5 ms. so:
60 * 32 * 4 * 0.5 = 3840ms (aprox. 4 seconds)

On the other hand, we have very obvious optimizations at hand: As the core of the problem is calculating a proof via a BLS signature based on slot + subcommittee index (Validator.Utils.get_sync_committee_selection_proof/4), the calculation is completely independent for each set of validator_index, slot, subcommittee_index. Given that, it's a low hanging fruit to use something like Task.async_stream to:

1. Parallelize computation per validator.
2. Inside of a validator, parallelize computation per slot/subcommittee.

As a followup question, is the all of the computation consumed in BLS.sign or is Misc.compute_signing_root a relevant portion of it? If it's BLS.sign, then there's no significant further change we can do. If it's compute_signing_root, however, that doesn't require the privkey, so we could do that once per subcommittee+slot instead of doing it once per {validator, slot subcommittee_index} tuple, although it's a bit of a bothersome refactor, so it's important to see if it's valuable at all.