Revert "Loads .pyd instead of .so in MemPool test for windows (#132749)"

This reverts commit 37ab0f33854fafdf9bf4f575260329ffcd960d13.

Reverted https://github.com/pytorch/pytorch/pull/132749 on behalf of https://github.com/syed-ahmed due to Seems like periodic is still failing: 7c79e89bc5 ([comment](https://github.com/pytorch/pytorch/pull/132749#issuecomment-2274041302))

test/test_cuda.py

@@ -4984,6 +4984,7 @@ class TestMemPool(TestCase):
         # increments the id
         self.assertTrue(abs(pool2[1] - pool1[1]) > 0)
 
+    @unittest.skipIf(IS_WINDOWS, "fails in periodic")
     def test_mempool_with_allocator(self):
         pool = torch.cuda.MemPool()
 
@@ -4993,11 +4994,9 @@ class TestMemPool(TestCase):
         from torch.utils.cpp_extension import load_inline
 
         dummy_allocator_source = """
-        #include <torch/extension.h>
         extern "C" {
-          // Note that windows needs __declspec(dllexport): https://stackoverflow.com/a/24575865
-          C10_EXPORT void* dummy_alloc(size_t size, int device, void* stream) { return nullptr; }
-          C10_EXPORT void dummy_free(void* ptr) { }
+          void* dummy_alloc(size_t size, int device, void* stream) { return nullptr; }
+          void dummy_free(void* ptr) { }
         }
         """
         dummy_allocator_libname = "dummy_allocator"
@@ -5009,7 +5008,7 @@ class TestMemPool(TestCase):
                 build_directory=tempdir,
             )
             allocator = torch.cuda.memory.CUDAPluggableAllocator(
-                dummy_allocator,
+                os.path.join(tempdir, f"{dummy_allocator_libname}.so"),
                 "dummy_alloc",
                 "dummy_free",
             )

torch/utils/cpp_extension.py

@@ -2143,7 +2143,6 @@ def _import_module_from_library(module_name, path, is_python_module):
         return module
     else:
         torch.ops.load_library(filepath)
-        return filepath
 
 
 def _write_ninja_file_to_build_library(path,