[xla:cpu] Switch XLA:CPU runtime to thunks interpreter

With this change XLA:CPU instead of compiling one LLVM function for the whole HLO module compiles separate functions for different fusions and runs them via the interpreter-like runtime.

This can change numerics because of slightly different LLVM IR and missed cross-fusion optimizations. If this breaks your tests, they likely have to relax numerical error tolerance.

Another potential issue is performance regressions for while loops with large number of iterations and small computation, as instead of compiling, we run such loops in interpreter. We plan to fix it in the future.

To disable thunks runtime set env variable: XLA_FLAGS=--xla_cpu_use_thunk_runtime=false.

PiperOrigin-RevId: 660520935

xla/debug_options_flags.cc

@@ -82,7 +82,7 @@ DebugOptions DefaultDebugOptionsIgnoringFlags() {
 #ifdef XLA_CPU_USE_ACL
   opts.set_xla_cpu_use_acl(true);
 #endif
-  opts.set_xla_cpu_use_thunk_runtime(false);
+  opts.set_xla_cpu_use_thunk_runtime(true);
   opts.set_xla_cpu_enable_concurrency_optimized_scheduler(false);
   opts.set_xla_cpu_prefer_vector_width(256);