[QNN EP] workaround for QNN validation bug for Tanh with uint16 quantized output (#23432)

### Description
- Skip QNN validation for Tanh with uint16 quantized output (workaround
for QNN validation bug).
- Re-enables unit test for Tanh with uint16 quantized output.

The [QNN
documentation](https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-50/HtpOpDefSupplement.html#tanh)
states that the output scale and offset for `ufixed_point_16` should be
(1/32768) and -32768, respectively. However, the QNN validator
incorrectly rejects these values. So, we skip validation for this
configuration of Tanh. Building an actual QNN graph with the correct
scale/offset still works.


### Motivation and Context
This QNN validation bug appeared in QNN SDK 2.28.0 and is still present
in QNN SDK 2.30.0. A previous PR disabled the corresponding unit test:
https://github.com/microsoft/onnxruntime/pull/22724/files#diff-57f590c6c548b073ba8cd8af6cf198799906f7059ea46b31cd33972ea9b01983R232

Author: adrianlizarraga

onnxruntime/core/providers/qnn/builder/opbuilder/simple_op_builder.cc

@@ -262,6 +262,22 @@ Status SimpleOpBuilder::ProcessAttributesAndOutputs(QnnModelWrapper& qnn_model_w
     if (node_unit.Domain() != kMSInternalNHWCDomain && (op_type == "DepthToSpace" || op_type == "SpaceToDepth" || op_type == "GridSample")) {
       return Status::OK();
     }
+
+#if QNN_API_VERSION_MAJOR >= 2 && QNN_API_VERSION_MINOR >= 21 && QNN_API_VERSION_MINOR <= 23
+    // Skip QNN validation for Tanh with uint16 (quantized) output.
+    // This gets around a Tanh QNN validation bug in QNN SDK 2.28.0 - 2.30.0.
+    // The QNN documentation states that the output scale and offset for ufixed_point_16 should be
+    // (1/32768) and -32768, respectively. However, the QNN validator incorrectly rejects these values.
+    if (op_type == "Tanh") {
+      TensorInfo output_info = {};
+      ORT_RETURN_IF_ERROR(qnn_model_wrapper.GetTensorInfo(node_unit.Outputs()[0], output_info));
+      if (output_info.qnn_data_type == QNN_DATATYPE_UFIXED_POINT_16) {
+        LOGS(logger, INFO) << "Skipping QNN validation for Tanh node '"
+                           << node_unit.Name() << "' with quantized unit16 output.";
+        return Status::OK();
+      }
+    }
+#endif
   }
 
   std::vector<std::string> param_tensor_names;

onnxruntime/core/providers/qnn/qnn_allocator.cc

@@ -60,7 +60,7 @@ size_t DivRoundUp(size_t a, size_t b) {  // TODO is there already a helper funct
 }
 
 bool IsAligned(const void* address, size_t alignment) {
-  assert((alignment & alignment - 1) == 0);  // alignment must be a power of two
+  assert((alignment & (alignment - 1)) == 0);  // alignment must be a power of two
   return (reinterpret_cast<uintptr_t>(address) & (alignment - 1)) == 0;
 }
 

onnxruntime/test/providers/qnn/simple_op_htp_test.cc

@@ -230,17 +230,19 @@ TEST_F(QnnHTPBackendTests, UnaryOp_Tanh) {
 }
 
 // disabled for QNN 2.28.0.241029 backendValidateOpConfig failed
-// still fails on QNN 2.28.2.
+// still fails on QNN 2.28.2 and QNN 2.30.0
 // QnnDsp <E> [4294967295] has incorrect Value -32768, expected equal to 0.
 // QnnDsp <V> validateNativeOps node_token_6:qti.aisw:Tanh htp op validator failed 3110
 // QnnDsp <V> registered validator failed => 3110
 // QnnDsp <E> QnnBackend_validateOpConfig failed 3110
 // QnnDsp <V> Wake up free backend (id: 1)'s thread(s)
 // QnnDsp <E> Failed to validate op node_token_6 with error 0xc26
 // Tests accuracy of 16-bit QDQ Tanh.
-TEST_F(QnnHTPBackendTests, DISABLED_UnaryOp_Tanh_U16) {
+//
+// We now skip QNN validation as a workaround for QNN SDK 2.28.0 to 2.30.0
+TEST_F(QnnHTPBackendTests, UnaryOp_Tanh_U16) {
   RunQDQOpTest<uint16_t>("Tanh",
-                         {TestInputDef<float>({1, 2, 3}, false, GetFloatDataInRange(-10.0f, 10.0f, 6))},
+                         {TestInputDef<float>({1, 2, 64}, false, GetFloatDataInRange(-10.0f, 10.0f, 128))},
                          {},
                          13,
                          ExpectedEPNodeAssignment::All,