CUDA/HIP: support argsort/top-k ncols>1024 on ROCm via hipCUB

[DeanoC]

Internal review PR. Stacked on #13 (base feat/topk-partial-sort-gfx1201), so the diff below is only the ncols>1024 enablement. Follows up the finding from #13 that N>1024 TOP_K/ARGSORT are unsupported on every ROCm build.

Problem

The GPU argsort/top-k uses a single-block bitonic sort (block_dims = ncols_pad), capped at 1024 threads/block. On CUDA, ncols>1024 is handled by the CUB device-sort path; on HIP that path is compiled out (GGML_CUDA_USE_CUB is CUDA-only), so supports_op returns ne[0] <= 1024 and larger sorts (full-vocab sampling, etc.) fall back to the CPU reference — 275 test-backend-ops cases on every ROCm build.

Fix

Enable the existing device-sort path on HIP via hipCUB (rocPRIM), which provides the same DeviceRadixSort / DeviceSegmentedSort / DeviceSegmentedRadixSort API:

• common.cuh: define GGML_CUDA_USE_HIPCUB on HIP.
• argsort.cu/.cuh: #include <hipcub/hipcub.hpp> and compile argsort_f32_i32_cuda_cub on HIP (no CCCL strided iterator → uses the existing init_offsets segment path).
• top-k.cu: on HIP, route ncols>1024 through the device sort + slice-first-k (mirrors the CUB branch); the fast partial-bitonic top-k still handles the common ncols ≤ 1024 decode/verify case unchanged.
• ggml-cuda.cu: supports_op returns true for TOP_K/ARGSORT when CUB or hipCUB is present.

Validation (test-backend-ops test)

op	arch	not-supported	fail
TOP_K	gfx1201 (R9700)	0 (was 275)	0
ARGSORT	gfx1201	0 (was 275)	0
TOP_K	gfx1151 (Strix Halo)	0 (was 275)	0
ARGSORT	gfx1151	0 (was 275)	0

The previously-unsupported ncols>1024 cases (up to ne=[262144]) now run on-GPU and are bit-for-bit vs the CPU reference on both archs. The ncols≤1024 partial-bitonic path (#13) is untouched.

Requires the ROCm hipcub + rocprim headers (ship with ROCm; present on our gfx1201 box and lucebox4 gfx1151 /opt/rocm).

[DeanoC]

Superseded by #15 — collapsed the partial-top-k win (#13) and the hipCUB ncols>1024 support (#14) into a single PR for review; they're complementary changes to the same kernels.