The implementation of approximated_rlp_size() and approximated_rlp_size_with_public() are sloppy

[fanlong]

[sparkmiw]

It is a bit tricky. It requires considerations on many aspects. Generally speaking, it is a performance/space/precision trade-off. First of all, when a node receives a block header and decodes it, we set the approximated_rlp_size as exactly rlp_size. However, when the node builds the block by itself, there are several options to do. One decision to make is when to do the rlp encode. It may do it immediately during the initialization to get the precise rlp_size, then when the block/header is propagated to peers later, it has to do the rlp encode again. The current solution is just to avoid doing the rlp encoding twice.