core: Reduce mutex contention in ActivityManager

 * applyOomAdjLocked in ActivityManager will call into native
   layer to adjust tasks in the cpuctl and memory cgroups. The global AM
   lock is held while this happens, and on a busy system can result in
   slowdowns.
 * The cpuctl cgroup is already handled in a sane way. Adjust the memory
   cgroup code to cache file descriptors to reduce overhead by an order
   of magnitude.
 * These cgroups are actually exactly the same in their content, in the
   future when remounting a cgroup fs is supported we can combine them.

Change-Id: If0043e651eebdf73075a96e6de618badd8b5935b