The transmit routine will stop interrupting and hang, causing the
tx_timeout routine to attempt to restart the device when the 32-bit
cur_tx counter wraps below dirty_tx.  If the device had called
netif_stop_queue it will never call netif_wake_queue in the interrupt
routine (at least on an IA32 system) due to 32-bit wrap around arithmetic.

On my IA32 system 'dirty_tx > lp->cur_tx - TX_RING_SIZE + 2' would
always evaluate to false when dirty and cur_tx were less than 15,
preventing netif_wake_queue to be called.

By starting dirty_tx and cur_tx at 0xfffffff0 (to reduce test time) I
found that once cur_tx wrapped to zero, that transmitted buffers would
never be unmapped or freed because 'while (dirty_tx < lp->cur_tx) {'
was not true.  cur_tx would keep incrementing (in start_xmit) but dirty_tx
would not (in pcnet32_interrupt), thus leaking skb's and pci map entries.
On PPC machines, the system would quickly run out of pci maps.

Fix tested on PPC and IA32.