Add 32-bit dispatch to pull_request workflow.

This is far more likely to break than 64-bit dispatch, so having it run per-PR makes sense.
64-bit dispatch is only tested in the nightly CI.

ci/matrix.yaml

@@ -28,6 +28,8 @@ workflows:
     - {jobs: ['build'], std: 'all', ctk: '12.5', cxx: 'nvhpc'}
     - {jobs: ['build'], std: 'all', cxx: ['gcc', 'clang'], cpu: 'arm64'}
     - {jobs: ['build'], std: 'all', cxx: ['gcc'], sm: '90a'}
+    # Test Thrust 32-bit-only dispatch here, since it's most likely to break. 64-bit-only is tested in nightly.
+    - {jobs: ['test_gpu'],  project: 'thrust', cmake_options: '-DTHRUST_DISPATCH_TYPE=Force32bit'}
     # default_projects: clang-cuda
     - {jobs: ['build'], std: 'all', cudacxx: 'clang', cxx: 'clang'}
     - {jobs: ['build'], project: 'libcudacxx', std: 'all', cudacxx: 'clang', cxx: 'clang', sm: '90'}
@@ -58,7 +60,7 @@ workflows:
     - {jobs: ['infra'], project: 'cccl', ctk: 'curr', cxx: ['gcc',   'clang']}
 
   nightly:
-    # Increased test coverage compared to nightlies:
+    # Increased test coverage compared to pull_request:
     - {jobs: ['test'],  std: 'all', cxx: ['gcc13', 'clang18', 'msvc2022']}
     - {jobs: ['test'],  project: 'cudax', ctk: ['12.0', 'curr'], std: 'all', cxx: ['gcc12']}
     - {jobs: ['test'],  project: 'cudax', ctk: ['12.0'        ], std: 'all', cxx: ['clang14']}