turbo Vulkan port: wire turbo KV cache into flash attention (Phase B)

Completes the turbo Vulkan port. turbo KV cache now runs end-to-end on Vulkan.

Phase B wiring:
- ggml-vulkan supports_op FLASH_ATTN_EXT: accept TURBO2/3/4_0 K/V (coopmat2-only,
  like Q1_0). fa_block_bytes already generic (ggml_type_size); TURBO_WHT op +
  SET_ROWS for turbo + cm2 dequantFuncTURBO3_0 survived earlier merges.
- common/arg.cpp: re-added turbo2/3/4 to kv_cache_types (-ctk/-ctv turbo3).

Status (AMD Radeon RX 9070 XT / RADV GFX1201):
- turbo3 KV cache RUNS via llama-bench (tg64 @ d2048 = 12.2 t/s).
- CORRECTNESS NOT VALIDATED on this GPU: RADV flash-attention is broken on
  RDNA4/GFX1201 for ALL types — test-backend-ops FLASH_ATTN_EXT 864/8453 pass,
  f16 KV perplexity = 9.4M with -fa vs 1.0092 without. Base (non-FA) Vulkan
  inference is correct (PPL 1.0092 == CPU); turbo weights validated. turbo KV
  correctness depends on FA, which this driver miscomputes for f16 too.
  -> Validate turbo KV on conformant-FA hardware (NVIDIA/CUDA or fixed RADV).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: retroheim

common/arg.cpp

@@ -397,6 +397,9 @@ const std::vector<ggml_type> kv_cache_types = {
     GGML_TYPE_IQ4_NL,
     GGML_TYPE_Q5_0,
     GGML_TYPE_Q5_1,
+    GGML_TYPE_TURBO2_0, // ik_llama port: TurboQuant 2-bit rotated KV
+    GGML_TYPE_TURBO3_0, // ik_llama port: TurboQuant 3-bit rotated KV
+    GGML_TYPE_TURBO4_0, // ik_llama port: TurboQuant 4-bit rotated KV
 };
 
 static ggml_type kv_cache_type_from_str(const std::string & s) {

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -16531,6 +16531,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     case GGML_TYPE_Q4_0:
                         return true;
                     case GGML_TYPE_Q1_0:
+                    case GGML_TYPE_TURBO2_0:
+                    case GGML_TYPE_TURBO3_0:
+                    case GGML_TYPE_TURBO4_0:
+                        // ik_llama port: turbo KV cache types — coopmat2 FA only
+                        // (flash_attn_cm2 decodes them via dequantFuncTURBO*).
                         return coopmat2;
                     default:
                         return false;