[rocm7.0_internal_testing] Prevent static initialization of at::cuda::warp_size() (#2293)

Fixes SWDEV-540240, SWDEV-540309, SWDEV-539989

```
...
```

80cca70
created a static global variable that used `at::cuda::warp_size()` to
initialize its value, which needs GPUs to be visible to query device
properties. However, GPUs are not present on CPU-only build systems.

Convert static variable into a static function, thus preventing static
initialization.

http://rocm-ci.amd.com/job/pyt_whl_docker_mainline/1461/artifact/build_artifacts.txt/*view*/

Ran microbenchmark to confirm basic functionality:
```
root@ubb4-rack-22:/var/lib/jenkins/pytorch-micro-benchmarking# python3 micro_benchmarking_pytorch.py --network resnet50
INFO: running forward and backward for warmup.
INFO: running the benchmark..
OK: finished running benchmark..
--------------------SUMMARY--------------------------
Microbenchmark for network : resnet50
Num devices: 1
Dtype: FP32
Mini batch size [img] : 64
Time per mini-batch : 0.10158218145370483
Throughput [img/sec] : 630.0317544289736=
```

Author: ethanwee1

aten/src/ATen/native/cuda/Embedding.cu

@@ -369,7 +369,7 @@ Tensor & embedding_renorm_cuda_(Tensor & self, const Tensor & indices,
 
     int warp_size = at::cuda::warp_size();
     TORCH_INTERNAL_ASSERT(num_threads() % warp_size == 0 &&
-                  num_threads() <= cuda_utils::kCUDABlockReduceMaxThreads,
+                  num_threads() <= cuda_utils::kCUDABlockReduceMaxThreads(),
                   "BlockReduceSum requires all warps be active");
     const int64_t *num_unique_indices_ptr = num_unique_indices.const_data_ptr<int64_t>();
     dim3 grid = unique_indices.numel();

aten/src/ATen/native/cuda/MultinomialKernel.cu

@@ -86,7 +86,7 @@ void renormRows(Tensor& t) {
   TORCH_CHECK(props != nullptr);
   int numSM = props->multiProcessorCount;
   const int64_t maxThreads = std::min(
-      props->maxThreadsPerBlock, cuda_utils::kCUDABlockReduceMaxThreads);
+      props->maxThreadsPerBlock, cuda_utils::kCUDABlockReduceMaxThreads());
 
   int warp_size = at::cuda::warp_size();
   dim3 grid(rows < numSM * 4 ? rows : numSM * 4);

aten/src/ATen/native/cuda/TensorModeKernel.cu

@@ -207,7 +207,7 @@ void handle_fused_mode(
   constexpr int num_threads = size / 2;
   int warp_size = at::cuda::warp_size();
   TORCH_INTERNAL_ASSERT(num_threads % warp_size == 0 &&
-                num_threads <= cuda_utils::kCUDABlockReduceMaxThreads, "");
+                num_threads <= cuda_utils::kCUDABlockReduceMaxThreads(), "");
   const auto memsize =
       (sizeof(scalar_t) * size) + (2 * size * sizeof(unsigned int));
   compute_mode<scalar_t, size>

aten/src/ATen/native/cuda/block_reduce.cuh

@@ -15,9 +15,13 @@ constexpr int kCUDABlockReduceNumThreads = 512;
 // ROCm NOTE: C10_WARP_SIZE should only be used inside device functions,
 // and kCUDABlockReduceMaxThreads is a host-side variable.
 #ifdef USE_ROCM
-static const int kCUDABlockReduceMaxThreads = at::cuda::warp_size() * at::cuda::warp_size();
+static int kCUDABlockReduceMaxThreads() {
+    return at::cuda::warp_size() * at::cuda::warp_size();
+}
 #else
-constexpr int kCUDABlockReduceMaxThreads = C10_WARP_SIZE * C10_WARP_SIZE;
+constexpr int kCUDABlockReduceMaxThreads() {
+    return C10_WARP_SIZE * C10_WARP_SIZE;
+}
 #endif
 
 // Sums `val` across all threads in a warp.