The current RPC mempool benchmark test tests a small mempool of
only 1-in-1-out txs. To more broadly capture different classes of perf
improvements (such as changes to UniValue), we can expand the mempool size
a bit and bloat the number of inputs and outputs.

In future diffs, perf improvements have become much more measurable.

Inspired by https://gitlab.com/bitcoin-cash-node/bitcoin-cash-node/-/merge_requests/102
though I don't see value in keeping the old test.