fix(cudf): Enable GPU execution for count(*), count(column), and count(NULL) (facebookincubator#16522)

Summary:
Enable the cuDF GPU code-path for all count aggregation variants:
count(*), count(column), count(constant), and count(NULL).

Previously, global count(*) fell back to CPU because its zero-column
intermediate representation loses row counts in cuDF.

- Support zero-column global count(*) on GPU by preserving row counts
  through FilterProject and CudfConversion.
- Classify count inputs (column, count-all, null-constant) to handle
  each variant correctly in both groupby and global reduce paths.
- Handle count(NULL) on GPU, returning 0 for all groups/globally.
- Non-count constant aggregates (e.g. sum(1)) fall back to CPU by
  design.

Tests:
- Re-enable previously disabled countStarGlobal test.
- Add parameterized tests for all count variants across single,
  partial+final, and partial+intermediate+final steps, for both
  global and group-by, with and without nulls.
- Add selection tests verifying GPU/CPU routing for zero-column
  count(*), count(NULL), and sum(1).

Fixes facebookincubator#16492

Pull Request resolved: facebookincubator#16522

Reviewed By: peterenescu

Differential Revision: D98979967

Pulled By: bikramSingh91

fbshipit-source-id: 2efa92c86fb30d25fcd785ff05f443d57e9a4c1a

Author: mattgara

velox/experimental/cudf/exec/CudfConversion.cpp

@@ -248,6 +248,19 @@ RowVectorPtr CudfToVelox::getOutput() {
     return nullptr;
   }
 
+  if (outputType_->size() == 0) {
+    // cuDF zero-column tables do not have a row count, so we sum the sizes
+    // of all CudfVectors in the inputs_, to maintain the logical count.
+    // This is necessary to ensure correct behavior for e.g. `count` operators.
+    vector_size_t totalSize = 0;
+    while (!inputs_.empty()) {
+      totalSize += inputs_.front()->size();
+      inputs_.pop_front();
+    }
+    finished_ = noMoreInput_ && inputs_.empty();
+    return BaseVector::create<RowVector>(outputType_, totalSize, pool());
+  }
+
   // Drain veloxBuffer_ (populated on a previous call) before consuming
   // more GPU inputs.
   if (!veloxBuffer_) {

velox/experimental/cudf/exec/CudfFilterProject.cpp

@@ -246,15 +246,19 @@ RowVectorPtr CudfFilterProject::getOutput() {
   VELOX_CHECK_NOT_NULL(cudfInput);
   auto stream = cudfInput->stream();
   auto inputTableColumns = cudfInput->release()->release();
+  auto outputSize = input_->size();
 
   if (hasFilter_) {
     filter(inputTableColumns, stream);
   }
+  if (!inputTableColumns.empty()) {
+    outputSize = inputTableColumns.front()->size();
+  }
   auto outputColumns = project(inputTableColumns, stream);
 
   auto outputTable = std::make_unique<cudf::table>(std::move(outputColumns));
   auto const numColumns = outputTable->num_columns();
-  auto const size = outputTable->num_rows();
+  auto const size = numColumns > 0 ? outputTable->num_rows() : outputSize;
   if (CudfConfig::getInstance().debugEnabled) {
     VLOG(1) << "cudfProject Output: " << size << " rows, " << numColumns
             << " columns";
@@ -263,7 +267,7 @@ RowVectorPtr CudfFilterProject::getOutput() {
   auto cudfOutput = std::make_shared<CudfVector>(
       input_->pool(), outputType_, size, std::move(outputTable), stream);
   input_.reset();
-  if (numColumns == 0 or size == 0) {
+  if (size == 0) {
     return nullptr;
   }
   return cudfOutput;