Always send PADDING_NEGOTIATE if middle supports it
We should define some kind of NULL machine for whatever hop is most common in our padding machine list, and negotiate that machine if no other machines apply to the current circuit. This machine shouldn't take up a slot or count as negotiated, tho, so we can still negotiate other machines at later points if the circuit purpose changes, etc.
Similarly, this NULL machine should (maybe) set should_negotiate_end and send a PADDING_NEGOTIATE at circuit close.
We need to do this so that there isn't an obvious PADDING_NEGOTIATE cell request/response pair with obvious timings that it went to the middle node (since the PADDING_NEGOTIATED response will come faster than all other responses on the circuit). See also legacy/trac#30092 (moved) for similar motivating reasoning.