Introduce heap balancer for the gRPC connection pool

[mrahs]

We observed increased latencies in builds that use the round-robin balancer. It is possible that the load is not uniformly spread across the pool, particularly around streams. This could cause client-side queuing on hot connections which may explain the increased latencies.

The heap-based balancer keeps track of the number of invocation and streaming requests per connection and selects the least used connection. One big caveat is that the pool is not aware of active connections. The number of stream requests is not necessarily the number of active streams on the connection. Without that knowledge, it is still possible for some connections to be used more than others. E.g. if long streaming requests accumulate on the same connection.