Add support for AMD GPU

[pxl-th]

This PR introduces AMD GPU support for gsplat.

Tested on:

• CPU: AMD Ryzen 7 5800X 8-Core @ 16x 4.594GHz
• GPU: Radeon RX 7900 XTX
• OS: Ubuntu 24.04
• ROCm: 6.4 (from official installation script).

Tested and confirmed working:

• 3DGS (simple & mcmc; packed & un-packed)
• 2DGS (packed & un-packed)
• 3DGUT

Relies on:

• AMD GPU enablement rahul-goel/fused-ssim#24
• AMD GPU enablement harry7557558/fused-bilagrid#6

Closes #771.
Closes #434.

• Move glm library out of the cuda/ directory to avoid hipifying it, which causes confusion during compilation.
• Since AMD GPU does not support cg::labeled_partition, simply avoid warp reductions on it and just do global memory writes directly. To reduce code duplication, introduce FOR_HIP variable that we use to determine whether to use labeled partition or just a placeholder to avoid compilation errors, e.g.:

#if FOR_HIP
auto warp_group_g = warp; // Not used, just here to not error in the if-statements.
#else
auto warp_group_g = cg::labeled_partition(warp, gid);
#endif

and in places where we'd do only one global atomic add (which should eliminate branching altogether):

if (FOR_HIP || warp_group_g.thread_rank() == 0) {
...
}

• Replace std::array.at with [] indexing to avoid bounds-checking which causes errors:

/usr/lib/gcc/x86_64-linux-gnu/13/../../../../include/c++/13/array:220:9: error: reference to __host__ function '__throw_out_of_range_fmt' in __host__ __device__ function
  220 |           std::__throw_out_of_range_fmt(__N("array::at: __n (which is %zu) "
      |                ^

• Since ROCm does not support cg::reduce, create equivalent warpSum reduction methods using shfl_down intrinsic & use it where we do reduction on the whole warp (i.e. tiled_partitiion in our case).
• Use respective NVCC flags depending on the GPU backend. E.g. -munsafe-fp-atomics is required to replace CAS-loop with fast hardware floating-point atomics that significantly improves the performance on AMD GPU.

[charyang-ai]

how is the performance comparied to Nvidia RTX card, like 4090?

[Spacefish]

https://github.com/Spacefish/gsplat-rdna/tree/rocm_7_2 haha did the same thing but on AMDs fork of gsplat.

Performance is ok, i guess 8x8 is impacted a littlebit as we have to use shared memory to accumulate 2 32 wide warps but other sizes are fine / work on 32 wide waves like a charm.