Bugfix: Using max where min was used by mistake to ensure a value is non-negative

[dhermes]

The original intent was to keep bytes at or above 0 but min([-n, 0]) == -n and more importantly min([n, 0]) = 0 so that bytes would be artificially dropped to 0 when drop() is called.

FWIW I was worried this was a data race issue, but I'm fairly certain it isn't. If in fact it turns out to be, we can add a no-op lock class on base.BasePolicy

import contextlib

@contextlib.contextmanager
def _no_op_lock():
    """No-op context manager."""
    yield

class BasePolicy(object):
    _LOCK_TYPE = _no_op_lock
    def __init__(self, ...):
        ...
        self._bytes_lock = self._LOCK_TYPE()

and then define _LOCK_TYPE = threading.Lock on our concrete implementation and use self._bytes_lock whenever modifying Policy().bytes.

Fixes #4516.