fatxpool: rotator cache size now depends on pool's limits

[michalkucharczyk]

Description

This PR modifies the hard-coded size of extrinsics cache within PoolRotator to be inline with pool limits.

The problem was, that due to small size (comparing to number of txs in single block) of hard coded size:

polkadot-sdk/substrate/client/transaction-pool/src/graph/rotator.rs

Line 34 in cdf107d

const EXPECTED_SIZE: usize = 2048;

excessive number of unnecessary verification were performed in prune_tags:

polkadot-sdk/substrate/client/transaction-pool/src/graph/pool.rs

Lines 369 to 370 in cdf107d

	let reverified_transactions =
	self.verify(at, pruned_transactions, CheckBannedBeforeVerify::Yes).await;

This was resulting in quite long durations of prune_tags execution time (which was ok for 6s, but becomes noticable for 2s blocks):

Pruning at HashAndNumber { number: 83, ... }. Resubmitting transactions: 6142, reverification took: 237.818955ms    
Pruning at HashAndNumber { number: 84, ... }. Resubmitting transactions: 5985, reverification took: 222.118218ms    
Pruning at HashAndNumber { number: 85, ... }. Resubmitting transactions: 5981, reverification took: 215.546847ms

The fix reduces the overhead:

Pruning at HashAndNumber { number: 92, ... }. Resubmitting transactions: 6325, reverification took: 14.728354ms    
Pruning at HashAndNumber { number: 93, ... }. Resubmitting transactions: 7030, reverification took: 23.973607ms    
Pruning at HashAndNumber { number: 94, ... }. Resubmitting transactions: 4465, reverification took: 9.532472ms    

Review Notes

I decided to leave the hardocded EXPECTED_SIZE for the legacy transaction pool. Removing verification of transactions during re-submission may negatively impact the behavior of the legacy (single-state) pool. As in long-term we probably want to deprecate old pool, I did not invest time to assess the impact of rotator change in behavior of the legacy pool.

[michalkucharczyk]

/cmd prdoc --bump minor --audience node_dev