Always compile the first invocation of a cluster

This is an optimistic guess that pays off for statically shaped TensorFlow
graphs (since they get the benefit of XLA right away without waiting for warmup)
and doesn't hurt much for dynamically shaped TensorFlow graphs (we "pay" at most
one cluster-compilation's worth of compile time).

Also rename CompileStats to ClusterCompileStats to make it more obvious that the
struct holds per-cluster information.

PiperOrigin-RevId: 218261029