Tests: revisit how stats are made available

[julen]

In #138 we made sure stats are made available to the test functions that require it by leveraging the refresh_stats fixture. The implementation might seem more like a workaround though.

We need to come up with a more robust solution that ensures predictability while not sacrificing speed.

Bringing here some ideas outlined in the original PR thread:

Improve TreeItem's testability

• Make TreeItem operations no-op by default in the test environment.
  This removes the chance for accidental poisoning. We can also think of alternative ways to make this class easier to test.
• Individual tests that require R/W access to stats can explicitly request so (via a custom marker, for instance), hence circumventing the default behavior.

Imitate rollback via code

• Redis has DUMP and RESTORE commands that might come handy. If we know that a specific test is going to alter certain keys, we can wrap it with the decorator that will backup keys before they are modified with the application code, and then restore the modified keys once the test is finished.
• This will allow us to actually test how the stats are affected during tests.