Make activity manager more robust in the face of app activity leaks.

This came up from bug #5601885: Memory increase (leak?) in system_server
Stingray MR1

This isn't *really* a leak in the system process -- it is a leak in an
application process that is causing the system process to keep around
a bunch of ActivityRecord objects longer than it should, until that app
process is ultimately killed.

Unfortunately these days leaking an ActivityRecord also often means
leaking a thumbnail, which is a big slab of memory.

So make the activity manager better about this, using a weak reference
from the handle the object has so we can still clean away most of the
state associated with the ActivityRecord even if the client side leaks
its own reference.

Change-Id: Idbab45e09749cdfb54899203da7981e7b3576e25