[CUDA] Add additional testcases for EraseUnwantedCUDAMatches.

[jlebar]

Summary

Specifically, this patch adds testcases for all three calls to
EraseUnwantedCUDAMatches. The addr-of-overloaded-fn test I accidentally
neutered in r264207, which moved much of
CodeGenCUDA/function-overload.cu into SemaCUDA/function-overload.cu.
The coverage from overloaded-delete test is new.

[tra]

LGTM.

cfe/trunk/test/SemaCUDA/addr-of-overloaded-fn.cu

+// expected-no-diagnostics
+
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fsyntax-only -verify %s
+// RUN: %clang_cc1 -triple nvptx64-nvidia-cuda -fsyntax-only -fcuda-is-device -verify %s
+
+#include "Inputs/cuda.h"
+
+__host__ void overload() {}
+__device__ void overload() {}
+
+__host__ __device__ void test_hd() {
+  // This should not be ambiguous -- we choose the host or the device overload
+  // depending on whether or not we're compiling for host or device.
+  void (*x)() = overload;
+}
+
+// These also shouldn't be ambiguous, but they're an easier test than the HD
+// function above.
+__host__ void test_host() {
+  void (*x)() = overload;
+}
+__device__ void test_device() {
+  void (*x)() = overload;
+}

cfe/trunk/test/SemaCUDA/overloaded-delete.cu

+// expected-no-diagnostics
+
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fsyntax-only -verify %s
+// RUN: %clang_cc1 -triple nvptx64-nvidia-cuda -fsyntax-only -fcuda-is-device -verify %s
+
+#include "Inputs/cuda.h"
+
+struct S {
+  __host__ static void operator delete(void*, size_t) {}
+  __device__ static void operator delete(void*, size_t) {}
+};
+
+__host__ __device__ void test(S* s) {
+  // This shouldn't be ambiguous -- we call the host overload in host mode and
+  // the device overload in device mode.
+  delete s;
+}
+
+__host__ void operator delete(void *ptr) {}
+__device__ void operator delete(void *ptr) {}
+
+__host__ __device__ void test_global_delete(int *ptr) {
+  // Again, there should be no ambiguity between which operator delete we call.
+  ::delete ptr;
+}