Switch to explicit atomic_batch() database handle

[carver]

What was wrong?

The old approach of batching all changes made while leveldb was locked was flawed in a multi-threading context: it permits unintended writes to be batched with the ones you want.

How was it fixed?

Return an explicit handle when making an atomic batch, so that all the batched changes you want to make are explicitly part of that atomic batch. eg~

atomic_db = AtomicDB()
with atomic_db.atomic_batch() as db:
    # changes are not immediately saved to the db, inside this context
    db[key] = val

    # changes are still locally visible even though they are not yet committed to the db
    assert db[key] == val

    # changes using the original db happen instantly, and are visible in the batch db
    atomic_db[key2] = val2
    assert db[key2] == val2

Cute Animal Picture

[carver]

We should expect performance gains once we start using write batches, as well.

[pipermerriam]

Maybe rename to decomission

[pipermerriam]

Rather than duplicating these here, do you think it makes sense to make this test suite generic or parametrized so that we can confirm feature parity between our in-memory databases and the leveldb ones?

[pipermerriam]

This name is odd, but am I correct that it's an internal API. If so, then I don't really care that much other than maybe prefixing with an _ to make that clear.