This query breaks down usage counts by relation and usage count, so you can see exactly how usage count distribution differs between tables:
The primary key index here isn't using that many buffers, but the ones it does have in the cache average a very high usage count. Meanwhile, the regular accounts data takes up much more room, but extremely little of it gets a usagecount greater than 1. This explains why the database considers pgbench_accounts_pkey as a popular relation, it should keep a large portion in memory, while pgbench_accounts is being treated as more transient data. This is really what you want. The ability to look up data in the accounts table requires having the primary key index, so you'd hope and expect as much of that index to be cached as possible.