We saw a case where a test (p2p_1p1c_network.py) called
raise_network_minfee(), which called fill_mempool() using node 0.
Then raise_network_minfee() returned, and the test called rescan_utxos(),
which called getrawmempool() using a different node (node 1) followed by
getrawtransaction() on each returned transaction, and the test asserted
because a transaction was not found.
This was caused by the timing window between the call to getrawmempool()
and fetching the individual transactions; the transactions were still
being propagated on the P2P network. During this window, a transaction
(returned by getrawmempool()) was evicted (the mempool is close to full
during this test), and did not exist in the mempool by the time it was
attempted to be fetched.
It might make more sense for rescan_utxos() to call sync_mempools() just
before calling getrawmempool(), but it can't because rescan_utxos() is
part of the MiniWallet class, which doesn't have access to test_framework
(but that could probably be changed).
Without this change, fill_mempool() may leave the mempool very close
to its memory size limit (-maxmempool). This can cause tests to
incorrectly fail when they submit another transaction expecting it
to succeed. Note that without this change, the same test that fails on
one platform may succeed on another, because their memory allocation
accounting algorithms (how they calculate memory usage, that is,
MallocUsage()) may be slightly different.
