Add support for 32KB multi_tensor_apply kernel arguments (#134373)

## Benchmark

On H100 SXM (HBM2e, 500W TDP), CUDA Toolkit=12.2, Driver Version=535.154.05, with [this script](https://gist.github.com/yifuwang/178c1f4bf951c5794ea79c04d90e44fa) (`torch._foreach_copy_`):

**Baseline**
```
https://interncache-all.fbcdn.net/manifold/perfetto-artifacts/tree/ui/index.html#!/?url=https://interncache-all.fbcdn.net/manifold/perfetto_internal_traces/tree/shared_trace/yifu_tmp0g_x4sys
device ms: 0.891, cpu ms: 7.200
memory bandwidth: 1457.727 GB/s
```

Single iteration trace:
<img width="1432" alt="image" src="https://github.com/user-attachments/assets/8ef54365-0265-4281-a0f0-d4c2f448300e">

**This PR**
```
https://interncache-all.fbcdn.net/manifold/perfetto-artifacts/tree/ui/index.html#!/?url=https://interncache-all.fbcdn.net/manifold/perfetto_internal_traces/tree/shared_trace/yifu_tmp3jqiugli
device ms: 0.683, cpu ms: 6.745
memory bandwidth: 1902.010 GB/s
```

Single iteration trace:
<img width="1074" alt="image" src="https://github.com/user-attachments/assets/e52acad1-d09b-492c-9611-6d69e339f3ac">

## Binary Size and Kernel Specialization
The binary size for `libtorch_cuda.so` increased 6MB (243MB -> 249MB).

```
// NOTE: [32KB kernel argument size support]
// 32KB kernel argument size support has three requirements:
// - CUDART_VERSION >= 12010
// - Driver version >= 530
// - GPU arch >= VOLTA
//
// Due to minor version compatibility, it possible for binaries built with
// CUDART_VERSION >= 12010 to run with driver version < 530. Since driver
// version can only be checked at runtime, if CUDART_VERSION >= 12010, we have
// to build both 4KB and 32KB kernels and determine the appropriate kernel to
// dispatch at runtime.
//
// - If CUDART_VERSION < 12010, only 4KB kernels will be instantiated.
//
// - If CUDART_VERSION >= 12010:
//   - Host code:
//     - We always instantiate the launching stub for both 4KB and 32KB kernels.
//   - Device code:
//     - If __CUDA_ARCH__ >= 700, we always instantiate both 4KB and 32KB
//     kernels.
//     - If __CUDA_ARCH__ < 700, it's not possible to even compile an empty
//     32KB kernel (formal parameter space overflowed). Thus, we only
//     instantiate a declaration for 32KB kernels. This is valid as long as the
//     declaration-only kernel is not launched.
//
// - At runtime, we dispatch to the 32KB kernel if driver version >= 530 and
// GPU arch >= VOLTA.
//
// - TODO(yifu): once there's a CUDART version that is not compatible with any
// driver version below 530, we can determine at compile time to not compile
// the kernels for 4KB kernel argument size.
//
// https://developer.nvidia.com/blog/cuda-12-1-supports-large-kernel-parameters/
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134373
Approved by: https://github.com/eqy, https://github.com/crcrpar, https://github.com/janeyx99

build_variables.bzl

@@ -1456,6 +1456,7 @@ aten_cuda_cu_source_list = [
     "aten/src/ATen/native/cuda/Equal.cpp",
     "aten/src/ATen/native/cuda/GridSampler.cpp",
     "aten/src/ATen/native/cuda/IndexKernel.cpp",
+    "aten/src/ATen/native/cuda/MultiTensorApply.cpp",
     "aten/src/ATen/native/cuda/ReduceOps.cpp",
     "aten/src/ATen/native/cuda/ScanKernels.cpp",
     "aten/src/ATen/native/cuda/Sort.cpp",