[Intel-gfx] Reduce interrupts-off time for execlists submission

[Chris Wilson]

Since we started doing direct submission to reduce the initial
submission stall, we split the submit tasklet into two: one part that
was wholly undercontrol of the tasklet, and the rest called from mixed
contexts. For the latter, we required additional extra spinlocks and
atomic handling to ensure concurrent operations, and worse, required
those sequences to use irq-off.

Instead of calling individual components of the submission process, we
can just call the whole tasklet and so replace the individual locking
with a known serialised calling context. That turns out to be far
simpler, and reduces the irqoff time required for the tasklet, improving
the parallelism of the hw. Slightly better latency, slightly higher
throughput, and slightly lower impact on system latency.
-Chris