[pytorch] fix CUDA_KERNEL_ASSERT macro for android build (#40151)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/40151
For debug android build it throws the following error:
```
In file included from src/pytorch/android/pytorch_android/src/main/cpp/pytorch_jni_common.cpp:9:
In file included from src/pytorch/android/pytorch_android/src/main/cpp/pytorch_jni_common.h:2:
In file included from ../../../../src/main/cpp/libtorch_include/armeabi-v7a/torch/csrc/api/include/torch/types.h:3:
In file included from ../../../../src/main/cpp/libtorch_include/armeabi-v7a/ATen/ATen.h:5:
In file included from ../../../../src/main/cpp/libtorch_include/armeabi-v7a/ATen/Context.h:4:
In file included from ../../../../src/main/cpp/libtorch_include/armeabi-v7a/ATen/Tensor.h:3:
In file included from ../../../../src/main/cpp/libtorch_include/armeabi-v7a/ATen/core/TensorBody.h:7:
In file included from ../../../../src/main/cpp/libtorch_include/armeabi-v7a/c10/core/Scalar.h:13:
../../../../src/main/cpp/libtorch_include/armeabi-v7a/c10/util/TypeCast.h:157:22: error: use of undeclared identifier '__assert_fail'
AT_FORALL_QINT_TYPES(DEFINE_UNCASTABLE)
^
```
Seems __assert_fail() isn't available on Android by default - in NDEBUG mode it forward declares the function and CI passes.
But CUDA_KERNEL_ASSERT() shouldn't be relevant for mobile build at all and we already bypass `__APPLE__` so the easiest fix is to add `__ANDROID__`.
Test Plan: Imported from OSS
Differential Revision: D22095562
Pulled By: ljk53
fbshipit-source-id: 793108a7bc64db161a0747761c0fbd70262e7d5a