Fix CUDA KV cache clear on non-host buffers

[double-thinker]

Note: This PR was assisted by AI but has been human-reviewed.

This fixes a CUDA crash when running with --gpu cuda.

The issue was caused by host-side memset/memmove on KV cache tensors (kv_self_k, kv_self_v) that can live in non-host backend buffers (CUDA/Vulkan/Metal). In that case, ggml_get_data() may not return host-safe memory, which can segfault.

Crash path (reproducible on GTX 1080 Ti, sm_61):

• voxtral_transcribe_from_audio()
• clear_kv_cache()
• memset(ggml_get_data(kv_self_*), ...)

gdb points to:

• src/voxtral.cpp:1082 (clear_kv_cache)

Fix

• In clear_kv_cache():
  • Detect whether KV tensors are host-backed via ggml_backend_buffer_is_host.
  • For non-host buffers, clear using backend-safe operations:
    • ggml_backend_buffer_clear(ctx->buf_persistent, 0) when available.
    • fallback to chunked ggml_backend_tensor_set(...) with zeroed data.
  • Keep direct memset only for host buffers.
• In kv_cache_shift_left():
  • Detect non-host KV buffers and avoid host memmove/memset.