Deadlock when adding watch callback in a callback

The current implementation of method `Etcd3Client.add_watch_callback` will trigger deadlock (block indefinitely) if it is invoked in a watch callback function.

For example:

```python
import time
import etcd3

etcd = etcd3.client(host='localhost')

def callback(event):
print 'Callback begins.'
etcd.add_watch_callback('/123', callback=callback)
print 'Callback done.'

callback(None) # Add the callback

print "Putting new key..."
etcd.put('/123', '123') # Trigger the added callback
print "New key was uploaded."

while True:
time.sleep(1) # Prevent the daemon watcher from being stopped
```

The second `"Callback done."` can never be printed if you run this code.

This problem is applicable to other methods which invoke `add_watch_callback`, including `watch`, `watch_once`, `watch_prefix` and `watch_prefix_once`.

Start by reading the implementation of Etcd3Client.add_watch_callback and the related watch, watch_once, watch_prefix, and watch_prefix_once methods named in the issue. Run the provided reproduction against a local etcd instance; done means adding a watch from inside a watch callback returns and the callback can continue without blocking indefinitely.

Written by the indexing model from the issue text.