This resolves an issue where getrawmempool() can race mempool
notification signals. Intuitively we use mempool.cs as a "read
lock" on the mempool with cs_main being the write lock, so holding
the read lock intermittently while doing write operations is
somewhat strange.
This also avoids the introduction of cs_main in getrawmempool()
which reviewers objected to in the previous fix in #12273

This backports core PR12368