more accurate row usage estimation in capacity checker

[lispc]

currently the algorithm is for each tx, we build an witness block, then calculate row usage of each subcircuit of this tx, then add the usage vector and find max inside the acc vector.

This algorithm will count some data redundantly. For example, if 2 tx shares same bytecode, then the len of bytecode circuit is counted twice. Overestimated.

We may impl a function "acc_witness_block += tx_witness_block"? Inside this function we do some basic de-duplicate and make more accurate estimation.

Pricinple: row estimation should be fast, a bit over estimation is allowd. Under estimation is seen as bug and should be avoided.