Feature/scm cuda backend

[Auc7us]

Add a CUDA backend for the SCM GPU path

The SCM GPU ray-cast and contact-force backend is HIP-only, an NVIDIA user needs ROCm
installed to use it. It now builds with either toolchain, and CUDA is what an NVIDIA machine
picks by default. Nothing is removed: HIP on ROCm and HIP through nvcc both still work.

Kernels are single-source. The two .cu files are compiled as CUDA or relabelled
LANGUAGE HIP by chrono_set_gpu_source_language(), the same arrangement Chrono::DEM and
Chrono::FSI::SPH already use. Device syntax is identical between the two languages, so only the
stream type and getLastError sit behind an #if.

Host bridges are duplicated, under terrain/gpu/cuda/ and terrain/gpu/hip/. They are
almost entirely runtime API calls, about thirty symbols, and Chrono has no compatibility layer
for those. Exactly one directory compiles per build; a banner at the top of each file says its
counterpart must be kept in step. Switch to a shim is possible if preferred.

CMake: REQUIRES HIP becomes REQUIRES CUDA_OR_HIP for the SCM GPU feature only.
CHRONO_GPU_BACKEND still selects explicitly; left at AUTO it resolves to CUDA on NVIDIA and
HIP on AMD.

No behavioural change, same kernels, same results. Verified.

Verification

Three toolchains, all clean, no source differences between them:

build	compiler	result
CUDA	nvcc 13.2, RTX 4080	builds, runs
HIP on NVIDIA	nvcc via CMAKE_HIP_PLATFORM=nvidia	builds, runs
HIP on AMD	ROCm 7.2 HIP-clang, gfx942	builds, runs, real gfx942 offload bundles

Benchmarks with the SCM visualization mesh off, 3 runs each, btest_VEH_hmmwvSCM and the Viper
SCM demo:

• CUDA vs HIP-through-nvcc on the 4080: CUDA 2.5–2.9% faster on HMMWV, parity on Viper.
• ROCm before vs after the refactor: every delta below the within-branch run-to-run spread.
• Viper step time, GPU ray cast vs CPU: 5.1x on the 4080, 6.4x on MI300X.
• Physics matches across vendors: wheel sinkage 0.05482–0.05484, final x −1.1401 to −1.1408 on
  both NVIDIA and AMD.