Keep objects in cache between tasks

[pcmoritz]

This PR changes the behavior so that if we get an object and let the object go out of scope, it will not unmap the object's memory immediately, only after a delay (i.e. after more objects went out of scope).
This is an important optimization, because unmap flushes the cpu cache. If tasks reuse the same objects, the objects sho ld stay in the cpu cache between task boundaries. This optimization is especially important for iterative workloads.

This illustrates the performance difference (it is ~2x slower without this PR).

A = np.random.rand(5000, 5000)
v = np.random.rand(5000)
A_id = ray.put(A)
v_id = ray.put(v)
a = time.time()
for i in range(100):
  A.dot(v)
b = time.time() - a
c = time.time()
for i in range(100):
  ray.get(A_id).dot(ray.get(v_id))
d = time.time() - c