drm/amdgpu: hold reference to fences in amdgpu_sa_bo_new (v2)

An arbitrary amount of time can pass between spin_unlock and
fence_wait_any_timeout, so we need to ensure that nobody frees the
fences from under us.

A stress test (rapidly starting and killing hundreds of glxgears
instances) ran into a deadlock in fence_wait_any_timeout after
about an hour, and this race condition appears to be a plausible
cause.

v2: agd: rebase on upstream

Signed-off-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Cc: stable@vger.kernel.org