Identify and lower library functions

• Add LibraryCall node for representing opaque library calls.
  • Integrate into collections analysis and GCM to properly track collection flow.
• Add rewrite-math-expressions pass to rewrite einsum reductions into LibraryCall nodes.
• Lower LibraryCall to functions in hercules_rt in RT backend.
• Add wrapper for cublasSgemm.
• Rewrite juno_matmul test on the GPU.

juno_samples/matmul/src/gpu.sch

@@ -12,9 +12,12 @@ fixpoint {
 fork-coalesce(*);
 infer-schedules(*);
 dce(*);
+rewrite(*);
+fixpoint {
+  simplify-cfg(*);
+  dce(*);
+}
 
-let out = auto-outline(*);
-gpu(out.matmul);
 ip-sroa(*);
 sroa(*);
 dce(*);