sched/fair: relax preferred cluster requirement

Although we steer the start_cpu towards preferred cluster, we reject
looking at other groups. This leads to unoptimal placements if the
task cannot be accommodated on the preferred cluster.

For example, for a task whose preferred cluster is the smallest cluster
but all the cpus in the smallest cluster cannot accommodate this task,
we skip looking at the higher clusters and end up returning -1 from
find_best_target() and end up choosing the prev_cpu or a fallback
scheme. All this when there could be a better cpu available in another
cluster.

Fix this by considering groups that do not belong to preferred
cluster. With this change preferred cluster will only be used to
influence the start cpu during placement.

Change-Id: I4ceea8eff04f6f015bedc4d689bfb3207fa0a5bc
Signed-off-by: Abhijeet Dharmapurikar <adharmap@codeaurora.org>