jit: skip expensive XLA graph passes when no clusters are compiled

[Copilot]

Enabling XLA on CPU causes 2–5% throughput regression even when zero subgraphs are compiled (and >5% with small subgraphs), due to unconditional overhead in graph optimization passes that runs regardless of whether any ops are actually marked for XLA compilation.

Changes

EncapsulateSubgraphsPass: early exit when no XLA clusters exist

The pass previously performed expensive setup work unconditionally:

• CPU Device creation via DeviceFactory
• ProcessFunctionLibraryRuntime construction
• Full graph copy inside EncapsulateSubgraphsInFunctions
• O(n) ref-variable analysis and per-node attribute writes

Now the pass checks for kXlaClusterAttr during the existing TPU-skip scan and returns early if no nodes were marked for compilation:

bool has_xla_clusters = false;
for (Node* n : (*options.graph)->nodes()) {
  if (n->type_string() == "TPUExecute" || ...) return absl::OkStatus();
  if (n->attrs().Find(kXlaClusterAttr)) has_xla_clusters = true;
}
if (!has_xla_clusters) return absl::OkStatus();  // skip all expensive setup

MarkForCompilation: O(1) early-exit on repeat invocations

After MarkForCompilationPassImpl::Run() completes, the source node (ID 0, always first in graph->nodes() iteration) is marked with kXlaAlreadyClustered. Previously, for graphs where no clusters were created, the existing early-exit scan found nothing and forced a full O(n) re-analysis on every subsequent invocation (e.g. from PartitionedCall re-running the optimization pipeline). The source-node marker converts that to O(1).

Test updates (encapsulate_subgraphs_pass_test.cc)

• RefVariablesMarked: updated to assert kXlaHasReferenceVarsAttr is not set when no clusters exist (attribute is only meaningful on XlaLaunch/_XlaCompile nodes)
• NoRefVarsNoAttr: adds explicit kXlaClusterAttr so the full encapsulation + ref-var analysis path is exercised
• Added NoXlaClustersEarlyExit: directly validates the early-exit optimization