This patch addresses two distinct races:

- Until now, the driver started the chip for Tx regardless of errors
  pending in the status register. Not good if an error occured while
  we were queueing packets -- the chip counter had not been reset,
  so Tx died. (We can't reliably get an interrupt for every error
  condition)

- The Rhine-II (when under load) frequently produces a Tx descriptor
  write-back race error. Failing to handle this means waiting for the
  netdev watchdog. Fixed.

  In addition, we must wait for the Tx engine to turn off on error
  conditions before we scavenge the descriptor entries. Failing to do
  so will typically lead to performance going down to about 10%: Burst,
  timeout, burst, timeout.. (again, with a Rhine-II under load).