perf: reduce payload build churn and cache loss during block assembly

[MatusKysel]

Miner: serialize payload builds and keep cached reads

• wrap build_payload/build_empty_payload in BuildResultWithCachedReads so cached reads are restored after each attempt
• gate concurrent payload builds with build_in_flight and keep best_payload inline instead of collecting a vector
• minor cleanup in builder: consume assembled_system_txs, avoid clones