Share "get_all_host_states" with concurrent requests

The scheduler cache-refresh operation (i.e. the function "get_all_host_states")
is very expensive. It needs to build a complete set of host states every time
when called by concurrent requests, causing more than 98% overhead in scheduler
service. However, the concurrent processing of "get_all_host_states" are
not necessary because they will return (almost) the same results based on the
same database queries and host manager caches.

This blueprint proposes an optimization to let concurrent request reuse the
return value of "get_all_host_states", so the host manager only need to build
host states once for all the incoming cache-refresh operations during the
building. This optimization can provide the freshest host states as before,
while minimize the database overhead from scheduler service, and accelerate
concurrent cache-refresh operations. It enables filter scheduler to support
much more compute nodes and heavier schedule workload.