candle-kernels 0.10.2 fails to build with CUDA 13.2 on Windows (CCCL requires /Zc:preprocessor)

## Describe the bug

`candle-kernels` fails to compile on Windows with CUDA 13.2 due to a CCCL preprocessor check.

The build aborts with:

```
fatal error C1189:
#error: MSVC/cl.exe with traditional preprocessor is used.
This may lead to unexpected compilation errors.
Please switch to the standard conforming preprocessor by passing /Zc:preprocessor to cl.exe.
You can define CCCL_IGNORE_MSVC_TRADITIONAL_PREPROCESSOR_WARNING to suppress this warning.
```

The error originates from:

```
include/cccl/cuda/std/__cccl/preprocessor.h
```

and is triggered while compiling `candle-kernels`.

---

## Environment

- OS: Windows 11 x64
- Rust: (stable)
- MSVC: 19.44.35225
- CUDA Toolkit: 13.2.78
- NVCC:

```
Cuda compilation tools, release 13.2, V13.2.78
```

- candle-core: 0.10.2
- candle-nn: 0.10.2
- candle-transformers: 0.10.2
- candle-kernels: 0.10.2

---

## Reproduction

Simply build a project that depends on Candle with CUDA enabled:

```bash
cargo build --release
```

The build fails while compiling `candle-kernels`.

---

## Build output

```
warning: candle-kernels@0.10.2: Compiling 11 of 11 PTX kernels

fatal error C1189:
#error: MSVC/cl.exe with traditional preprocessor is used.
Please switch to the standard conforming preprocessor by passing /Zc:preprocessor to cl.exe.

Error: CompilationFailed {
path: "src\\reduce.cu",
message: "nvcc error:"
}
```

---

## Additional information

MSVC itself is recent and supports `/Zc:preprocessor`.

```
Microsoft (R) C/C++ Optimizing Compiler Version 19.44.35225 for x64
```

This appears to be caused by `nvcc` invoking `cl.exe` without forwarding `/Zc:preprocessor`.

CUDA 13.2's CCCL headers now require the standard-conforming preprocessor and emit a hard error otherwise.

Would it make sense for `candle-kernels`' build script to pass:

```
-Xcompiler=/Zc:preprocessor
```

when compiling CUDA kernels on Windows, or is CUDA 13.2 currently unsupported?

Thanks!

Start by running `cargo build --release` with CUDA enabled and trace how candle-kernels invokes nvcc for `src\reduce.cu` on Windows. Inspect the path reaching `include/cccl/cuda/std/__cccl/preprocessor.h` and verify whether compiler flags are forwarded to cl.exe. Done means candle-kernels builds successfully with CUDA 13.2 on Windows or documents that version as unsupported.

Written by the indexing model from the issue text.