Pjpersist

------------------

- Fix `DBRef` comparisons to return valid results instead
of failing hard when comparing to `None` and not `DBRef` instances.
(`__neq__` was unused because `__ne__` is the right method and `__ne__` anyway
delegates to `__eq__`)

------------------

- Fix imports in `testing.py`

- Declare Python 3.9 compatibility

- Small improvement in `pjpersist.zope.container.PJContainer._load_one`
Get the local cache just once, since the `_cache` property became
a bit more expensive.

- Added `sqlbuilder.ILIKE` -- case-insensitive pair of LIKE

------------------

- Backwards incompatible change: PJDataManager now accepts a pool instead
of connection object. PJDataManager will get the connection from the pool
when joining the transaction, and return it back when transaction
completes (aborts or commits). This allows for more flexible connection
management. The connection pool must implement IPJConnectionPool interface
(it is compatible with psycopg2.pool).

- `IPJDataManager.begin()` is renamed to `setTransactionOptions()`

- Errors executing SQL statements now doom the entire transaction,
causing `transaction.interfaces.DoomedTransaction` exception on
any attempts to commit it. A failed transaction must be aborted.

------------------

- Fixed persisting tuple keyed dicts. Persisting such objects worked,
but reading failed.

------------------

- Drop Python 2.7 and 3.6 support, add 3.8.

- Remove buildout support.

- Support for nested flushing. In complex use cases it can happen that during
serialization of an object, a query is made to look up another object. That
in turn causes a flush, resulting in a flush inside a flush. The `flush()`
method did not expect that behavior and failed if the inner flush would
flush objects that the outer flush had already handled.

------------------

- Optimization: do not dig in `data` when we have a native field for
`_pj_mapping_key`, should allow creating indexes for lookup