[ath] Hopefully recover better-er upon RX restart on AR9380.

This is all very long-standing bug stuff that is touchy and still poorly
documented. Ok, here goes.

The basic bug:

• deleting a VAP causes the RX path (and TX path too) to be restarted without a full chip reset, which causes RX hangs on the AR9380 and later. (ie, the ones with the newer DMA engine.)

The basic fix:

• do an RX flush when stopping RX in ath_vap_delete() to match what happens when RX is stopped elsewhere. This ensures any pending frames are completed and we restart at the right spot; it also ensures we don't push new RX buffers into the hardware if we're stopping receive.

The other issues I found:

• Don't bother checking the RX packet ring in the deferred read taskqueue; that's specifically supposed to be for completing frames rather than just yanking them off the receive ring.

• Cancel/drain any pending deferred read taskqueue. This isn't done inside any locks so we should be super careful here. This stops the hardware being reprogrammed at the same time in another thread/CPU whilst we're stopping RX.

• .. (yes, this should be better serialised, but that's for another day. maybe.)

• Add more debugging to trace what's going on here.

And the fun bit:

• Reinitialise the RX FIFO ONLY if we've been reset or stopped, rather than just reset. I noticed that after all the above was done I was STILL seeing RXEOL. RXEOL isn't enabled on the AR9380 so I'd only see it if I was sending TX frames (ie a ping where it'd be transmitted but never received) so I was not being spammed by RXEOL. So, as long as stuff is stopped, restart it.

This seems to be doing the right thing in both AP and STA modes.

What I should do next, if I ever get time:

• as I said above, serialise the receive stop/start to include taskqueues
• monitor RXEOL on the AR9380 and I keep seeing it spammed / lockups, just go do a full chip reset to get things back on track. It sucks, but it is better than nothing.

Tested:

• AR9380 AP/STA mode, adding/deleting a hostap VAP to trigger the TX/RX queue stop/start; whilst also running an iperf through it. Lots of times. Lots. Of.. Times.