Compute packed leaf values on demand (Option 1)

[macladson]

Addresses #10

Keeps the hash field and computes the values field on demand.

Currently a WIP awaiting comparisons to keeping just the values field and computing hash on demand.

[michaelsproul]

Looking good and getting close to ready I think. I just found one more small optimisation related to the removal of the hash from PackedLeaf::update.

[michaelsproul]

I did some CPU benchmarking of this branch vs main vs #18.

The results are a bit frustrating.

Slot processing and block processing are both around 1% faster with this branch, but tree hashing seems to be a bit slower (~10%). Full data here: https://docs.google.com/spreadsheets/d/1IbgU8wvLeC1XzDb3tQU-QQM0htHlyjoYlbVnG-kH8Dw/edit?usp=sharing.

The Lighthouse commit used was sigp/lighthouse@bba1526.

The command used was lcli, compiled as so:

cargo build --profile maxperf --bin lcli --features modern,jemalloc

The benchmarking command used was:

RUST_LOG=debug ./target/maxperf/lcli transition-blocks --pre-state-path ~/Programming/lighthouse/state_6300863.ssz --block-path ~/Programming/lighthouse/block_6300864.ssz --post-state-output-path post_state.ssz --exclude-cache-builds --runs 200 2> stock2.log

The log files were then parsed using rg and sed to produce the data in the table.

I had a go at making some more optimisations in this commit, which is dedup3 in the data: b445f24. It seemed to make things worse.

--

If we can demonstrate a clear memory benefit I think we should still go ahead with this change. That might just be a bit trickier to establish. It really seems like this branch should be better.