[Intel-gfx] No premature unpinning

[Chris Wilson]

Ok, we had to take a little bit more of a detour. An issue BAT detected
was that if we interrupted building the request, then due to the
retire/free reorganisation we leaked a reference on the context. That
required a little bit more reshuffling of request_alloc, and I'm not
entirely happen with the intermediate state we leave it in (more changes
are required to make the divergence in request handling between legacy
and execlists disappear).
-Chris