fix(hip): dflash compatibility aliases + TQ3_0 FA guard

[smpurkis]

Note: This PR was vibe-coded with AI assistance (Claude Code). The fixes are mechanical (alias additions + HIP-gated guards) — please review accordingly.

Summary

• Three small HIP/ROCm compatibility fixes needed to build the luce-dflash branch under -DGGML_HIP=ON for downstream dflash builds (Strix Halo / gfx1151, ROCm 7.2 / clang 21).
• No behavior change for CUDA builds — all changes are HIP-gated or alias-only.

Changes

1. ggml/src/ggml-cuda/vendors/hip.h — add four missing CUDA→HIP symbol aliases that newer dflash code paths depend on:
   • cublasSgemmStridedBatched → hipblasSgemmStridedBatched
   • cudaStreamCaptureStatus → hipStreamCaptureStatus
   • cudaStreamCaptureStatusNone → hipStreamCaptureStatusNone
   • cudaStreamIsCapturing → hipStreamIsCapturing
2. ggml/src/ggml-cuda/gated_delta_net.cu — guard the <cuda_fp16.h> include under #ifndef GGML_USE_HIP (HIP path picks up the type via hip_fp16.h through the runtime header).
3. ggml/src/ggml-cuda/fattn.cu — wrap the TQ3_0 FATTN_VEC_CASES_ALL_D template instantiations under #ifndef GGML_USE_HIP. The TQ3_0 vec_dot kernel isn't ported to HIP yet, so instantiating these on HIP fails to link.

Why

Without these, a HIP build of luce-dflash fails at ggml-hip with errors like:

• 'cuda_fp16.h' file not found (gated_delta_net.cu)
• use of undeclared identifier 'cublasSgemmStridedBatched' / 'cudaStreamCaptureStatus' (fattn-chunked.cu)
• TQ3_0 vec-FA template instantiation failure (fattn.cu)

Test plan

• Builds cleanly on Strix Halo (gfx1151) with ROCm 7.2 + -DGGML_HIP=ON -DCMAKE_HIP_ARCHITECTURES=gfx1151
• Used end-to-end as the submodule for the lucebox-hub HIP/ROCm support PR — Add HIP/ROCm support for Strix Halo (gfx1151) lucebox-hub#119
• Existing CUDA paths untouched (no instantiations removed; only HIP-gated)