Fix bugs with 32 bit types being passed to gpuScan (#24600)

There were failures in `gpu-sort-performance-explorer` caused by in
in-house radix sort using the `gpuScan` which was untested for
`uint(32)` types.
Upon testing it was revealed that passing any type which is not a 64 bit
type caused an error as seen in
#24602

This PR fixes `gpuScan` to work with no 64 bit types by using the
workaround described in issue above. Why this works is not known yet :)

While there, we also expand the `gpuScan` testing and cleanup some code.

[Reviewed by @vasslitvinov]

modules/standard/GPU.chpl

@@ -795,7 +795,7 @@ module GPU
 
   // This function requires that startIdx and endIdx are within the bounds of the array
   // it checks that only if boundsChecking is true (i.e. NOT with --fast or --no-checks)
-  private proc serialScan(ref arr : [] ?t, startIdx = arr.domain.low, endIdx = arr.domain.high) {
+  private inline proc serialScan(ref arr : [] ?t, startIdx = arr.domain.low, endIdx = arr.domain.high) {
     // Convert this count array into a prefix sum
     // This is the same as the count array, but each element is the sum of all previous elements
     // This is an exclusive scan

test/gpu/native/studies/sort/cuda12_misaligned.bad

@@ -0,0 +1 @@
+internal error: gpu-nvidia.c:292: Error calling CUDA function: misaligned address (Code: 716)

test/gpu/native/studies/sort/cuda12_misaligned.chpl

@@ -0,0 +1,17 @@
+on here.gpus[0] {
+  var a: [1..10] uint(32);
+  @assertOnGpu
+  foreach chunk in 0..1 {
+    serialScan(a);   // Exclusive scan in serial
+  }
+}
+
+ proc serialScan(ref arr : [] uint(32)) {
+  // Calculate the prefix sum
+  var sum : uint(32);
+  var temp : uint(32) = arr[1];
+  arr[1] = sum;
+  sum += temp;
+
+}
+

test/gpu/native/studies/sort/cuda12_misaligned.future

@@ -0,0 +1 @@
+# issue: 

test/gpu/native/studies/sort/gpu-sort-performance-explorer.chpl

@@ -15,24 +15,20 @@ config const chunkSize = 6250;
 config const bitsAtATime = 8;
 
 config const inputDataScheme = 1;
-
-config const parallel = true;
-
-config param reverse = false;
 config type eltType = uint(32);
 
 config const seed = NPBRandom.oddTimeSeed();
 
 var methods = ["default", "gpuCub", "gpuRadix"];
 
-proc testsort(ref input, method) {
+proc testsort(ref gpuArr, method) {
 
   if method == "gpuCub" {
-      GPU.gpuExternSort(input);
+      GPU.gpuExternSort(gpuArr);
   } else if method == "gpuRadix" {
-    GPU.parallelRadixSort(input, bitsAtATime, chunkSize, false);
+    GPU.parallelRadixSort(gpuArr, bitsAtATime, chunkSize);
   } else if method == "default" {
-    GPU.gpuSort(input);
+    GPU.gpuSort(gpuArr);
   } else {
     halt("Unknown sorting method " + method);
   }
@@ -145,11 +141,11 @@ proc testsize(size:int) {
     for i in 1..ntrials {
       input = makeInput(array);
       on here.gpus[0]{
-        var arr = input; // Copy to gpu
+        var gpuArr = input; // Copy to gpu
         t.start();
-        testsort(arr, m);
+        testsort(gpuArr, m);
         t.stop();
-        input = arr; // Copy back to cpu
+        input = gpuArr; // Copy back to cpu
       }
     }
     var mibs = mibibytes * ntrials / t.elapsed();

test/gpu/native/studies/sort/scanTest.chpl

@@ -28,8 +28,15 @@ if multiDimArray {
 var timer: Time.stopwatch;
 
 testType(uint);
+testType(uint(32));
+testType(uint(16));
+testType(uint(8));
 testType(int);
+testType(int(32));
+testType(int(16));
+testType(int(8));
 testType(real);
+// testType(real(32)); // Causes overflow (I think?)
 
 proc testType(type t) {
   var cpuArr: [low..#arrSize] t;