This improves concurrency by not locking CTxMemPool every time CBlockPolicyEstimator is used (although as said this is incomplete and the concurrency improvements can be continued), so it may have a positive impact on performance. But I haven't done any benchmark or test.

My original plan was to do this at once while encapsulating the global minRelayTxFee, but relay policy encapsulation is not welcomed at this point and now it would be more disruptive to completely decouple CTxMemPool from CBlockPolicyEstimator than it used to be when first coded this many months ago.
We can take the first steps instead of waiting to do it at once, and that would hopefully prevent the code from evolving to something where is even harder to make this happen.
Although @morcos thinks that CBlockPolicyEstimator should depend on CTxMemPool and I disagree on that, we both agree that CTxMemPool should not depend on CBlockPolicyEstimator. This PR reduces CTxMemPool's dependency on CBlockPolicyEstimator.