[BE] Make torch.cuda.has_magma a build time check (pytorch#116299)

Perhaps originally one needed to query about GPU capability, but right now it's a simple check for a build time flag: https://github.com/pytorch/pytorch/blob/52f0457d7dc8e1ee6e25b6c97d11d3070d11263b/aten/src/ATen/cuda/detail/CUDAHooks.cpp#L165-L171

Alternative, to avoid `at::hasMAGMA()` call  one can implement it as follows:
```cpp
  const auto use_magma = caffe2::GetBuildOptions().at("USE_MAGMA");
  return PyBool_FromLong(use_magma == "1");
```

Make this check very similar to `_has_mkldnn`
https://github.com/pytorch/pytorch/blob/0978482afa69118da1a986a4edec3acea01d2c6d/torch/csrc/Module.cpp#L1793-L1794

Test plan:
 Run `lldb -- python3 -c "import torch;print(torch.cuda.has_magma)"` and make sure it returns True and that `cuInit` is not called

Pull Request resolved: pytorch#116299
Approved by: https://github.com/seemethere, https://github.com/albanD

Author: malfet

torch/_C/__init__.pyi.in

@@ -1208,6 +1208,7 @@ has_mkl: _bool
 _has_mps: _bool
 has_lapack: _bool
 _has_cuda: _bool
+_has_magma: _bool
 _has_mkldnn: _bool
 _has_cudnn: _bool
 has_spectral: _bool

torch/csrc/Module.cpp

@@ -1789,6 +1789,8 @@ Call this whenever a new thread is created in order to propagate values from
 #endif
 
   ASSERT_TRUE(set_module_attr("_has_cuda", has_cuda));
+  ASSERT_TRUE(
+      set_module_attr("_has_magma", at::hasMAGMA() ? Py_True : Py_False));
   ASSERT_TRUE(set_module_attr("_has_mps", has_mps));
   ASSERT_TRUE(
       set_module_attr("_has_mkldnn", at::hasMKLDNN() ? Py_True : Py_False));

torch/csrc/cuda/Module.cpp

@@ -1285,8 +1285,6 @@ static PyObject* THCPModule_initExtension(PyObject* self, PyObject* noargs) {
     }
   };
 
-  set_module_attr("has_magma", at::hasMAGMA() ? Py_True : Py_False);
-
   auto num_gpus = c10::cuda::device_count();
   auto default_cuda_generators = PyTuple_New(static_cast<Py_ssize_t>(num_gpus));
   for (const auto i : c10::irange(num_gpus)) {

torch/cuda/__init__.py

@@ -111,8 +111,8 @@ def _maybe_exchange_device(device: int) -> int:
 
 
 has_half: bool = True
-# Global variables dynamically populated by native code
-has_magma: bool = False
+has_magma: bool = torch._C._has_magma
+
 default_generators: Tuple[torch._C.Generator] = ()  # type: ignore[assignment]