Port CUDA Kernel to HIP

Port CUDA kernels to HIP using hipify for initial API mapping, then apply manual fixes for wavefront width differences and memory model adaptations to maintain performance intent on AMD ROCm devices.

AMD HIP uses 64-wide wavefronts compared to CUDA's 32-wide warps. HIP porting requires adapting kernel indexing and synchronization logic to handle this wavefront width difference correctly.

Hipify provides initial CUDA to HIP API mappings, but manual fixes are required across the codebase to handle semantic differences, memory model adaptations, and wavefront width differences.

Validate HIP kernel correctness by compiling, testing, and verifying numerical equivalence against the CUDA reference output on an AMD device to ensure functional parity after migration.

Porting CUDA to HIP requires memory model adaptations to handle semantic differences between NVIDIA and AMD hardware, ensuring correct memory access patterns and performance intent preservation.