Add missing __syncwarp() in reduce kernel + CUDA 13.0 build fix

[cuzelac]

Summary

Two fixes discovered while debugging Trellis2 on NVIDIA Blackwell (RTX 5090, sm_120):

1. Missing __syncwarp() in warp-level reduction

reduce_code_cuda_kernel in neighbor_map.cu performs a warp-level reduction over shared memory without __syncwarp() between iterations. Each iteration reads buf[threadIdx.x + cur_len] after the prior iteration wrote to buf[threadIdx.x]. Without a warp barrier, there is no guarantee the write is visible to other threads before the next read.

While current NVIDIA hardware executes warps in lockstep, this is undefined behavior per the CUDA programming model and may break on future architectures.

2. -allow-unsupported-compiler for CUDA 13.0

CUDA 13.0 with MSVC 2025 (Visual Studio 18) fails to compile without this flag, as nvcc only officially supports MSVC 2019-2022.

Testing

• Built and tested on RTX 5090 (sm_120), PyTorch 2.10.0+cu130, CUDA 13.0, Windows
• Full Trellis2 image-to-3D pipeline including sparse convolution operations
• Multiple successful runs

[cuzelac]

This may also address:

• cuda 13.0 support #8 (CUDA 13.0 build support) — the -allow-unsupported-compiler flag in this PR enables building on CUDA 13.0
• Silent failure in RTX 5080 #13 (silent failure on RTX 5080) — the __syncwarp() fix or the related CuMesh stream fix may resolve this

[JeffreyXiang]

Thanks! Merged

[cuzelac]

Happy to contribute - thanks for your work on this!