[Intel-gfx] Remove HW semaphore synchronisation for gen7

[Chris Wilson]

So a global per-engine execution queue is planned to go the way of the
dodo. One consequence for this is that HW semaphores will no longer be
viable for gen7. However, this is a mixed blessing as it gives us a
strong impetus to remove them and take another step towards implementing
request reordering. (Now we just need move our ring into per-context
batch buffer and use it as a trampoline for each request. Almost there.)

Using a roundtrip via the cpu is much slower than using the gpu to kick
the other engine, so by means of compensation optimise the interrupt
handling latency and improve request generation. Scratch the improve
interrupt latency since that broke display interrupts in the process.
Sigh it was a 50% improvement in latency for GEM. :(
-Chris