Use auto-incremented stream registers to signal dispatcher that workers are done.

[jbaumanTT]

If we use auto-incrementing stream registers, we can continue to use posted transactions (in this case inline writes) on Blackhole to do the updates, since they're not affected by the same bug as posted (or inline) transactions to L1.

Since many cores are writing to the same destination, a stream register has better latency and can better handle the traffic than L1 could, reducing the done->GO latency by around 70 cycles.