Currently, plasma evicts objects according a global LRU queue. In Ray, this often causes memory-intensive workloads to fail unpredictably, since a client that creates objects at a high rate can evict objects created by clients at lower rates. This is despite the fact that the true working set of both clients may be quite small.

This PR adds

• support for configuring per-client LRU eviction queues at runtime
• improved debugging of memory usage of clients

Details on the implementation:

QuotaAwarePolicy extends the basic eviction policy to implement per-client memory quotas. This effectively gives each client its own LRU queue, which caps its memory usage and protects this memory from being evicted by other clients.

The quotas are enforced when objects are first created, by evicting the necessary number of objects from the client's own LRU queue to cap its memory usage. Once that is done, allocation is handled by the normal eviction policy. This may result in the eviction of objects from the global LRU queue, if not enough memory can be allocated even after the evictions from the client's own LRU queue.

Some special cases:

• When a pinned object is "evicted" from a per-client queue, it is instead transferred into the global LRU queue.
• When a client disconnects, its LRU queue is merged into the head of the global LRU queue.

cc @pcmoritz