The slow fuzz seed described in #27799 was just slower than expected,
not an endless loop. Ensuring that every anscestor is only processed
once speeds up the termination of the graph traversal.
Fixes #27799
Rewrite the same algo instead of reusing BlockAssembler because we have
a few extra requirements that would make the changes invasive and
difficult to review:
- Only operate on the relevant transactions rather than full mempool
- Remove transactions that will be replaced so they can't bump their ancestors
- Don't hold mempool lock outside of the constructor
- Skip things like max block weight and IsFinalTx
- Additionally calculate fees to bump remaining ancestor packages to target feerate
Co-authored-by: Murch <murch@murch.one>
