perf(engine): pre-populate keccak cache before parallel tasks begin

[i-m-aditya]

Ref #20434

[mediocregopher]

@i-m-aditya the idea here is good but the current implementation needs work. By calling tx_iterator_for_payload inside the warming logic you're decoding/recovering txs, which is also done later on in the validation, resulting in duplicate work. If you could de-duplicate this work then this change will likely be a net win.