Auto-cluster TensorFlow's resource operations when legal.

This CL teaches auto-clustering to cluster resource variable operations when
proved legal by a data-flow analysis.  It is organized as follows:

 - The core data flow analysis lives in jit/resource_operation_safety_analysis.
   The output of this analysis is a set of pairs of nodes where two nodes that
   belong to the same pair should not be put into the same cluster.

 - jit/xla_cluster_util now has a routine that incorporates the results of the
   data flow analysis into the cycle detection graph such that auto-clustering
   automatically avoids putting incompatible nodes in the same cluster.

 - jit/mark_for_compilation_pass and jit/xla_fusion_optimizer have some plumbing
   changes to make all of this work.  Changes to jit/xla_fusion_optimizer are a
   bit simpler because, IIUC, it does not have to deal with the XLA_{CPU|GPU}
   devices.

PiperOrigin-RevId: 210006397