ggml: add Vulkan and SYCL backends for TQ3_0 KV cache quantization

[metalchef1]

Add hardware-accelerated TQ3_0 support for Vulkan and SYCL backends, enabling TQ3_0 KV cache on Intel Arc GPU with 58% VRAM savings vs q8_0.

Vulkan backend:

• types.glsl: add block_tq3_0_packed16 struct and A_TYPE_PACKED16 macro
• dequant_funcs.glsl: add dequantize/dequantize4/get_dm for TQ3_0 using 32-point Walsh-Hadamard Transform with Max-Lloyd centroids
• copy_to_quant.comp: TQ3_0 quantization path (F32 -> TQ3_0)
• dequant_tq3_0.comp: standalone dequantize shader (TQ3_0 -> F32)
• flash_attn_base.glsl: TQ3_0 dequantize4 for flash attention K/V path
• vulkan-shaders-gen.cpp: compile TQ3_0 shaders (cpy, dequant, mul_mat_vec, get_rows, flash_attn scalar path)
• ggml-vulkan.cpp: pipeline creation for cpy/dequant/mul_mat_vec/flash_attn, supports_op entries for CPY/DUP/CONT/MUL_MAT/FLASH_ATTN_EXT

SYCL backend:

• dequantize.hpp: TQ3_0 WHT dequantize kernel
• dmmv.cpp: dot-matrix-vector kernel for TQ3_0 x F32
• convert.cpp: F32 -> TQ3_0 and TQ3_0 -> F32 copy kernels
• ggml-sycl.cpp: register TQ3_0 pipelines

Shared:

• ggml-common.h: add QR_TQ3_0 = 1 macro
• llama-context.cpp: enable flash_attn for TQ3_0 KV cache (was force-disabled)

Tested on Intel Arc Pro B60 (24GB) with Qwen3.6-35B-A3B-IQ4_XS:
ctx=32768, K+V q8_0: 340 MiB
ctx=32768, K+V tq3_0: 140 MiB (58.8% reduction, correct outputs verified)

TQ3_0 uses flash attention (FLASH_ATTN_EXT) since the K cache is accessed non-contiguously. Requires --cache-type-v tq3_0 alongside --cache-type-k tq3_0.

Overview

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: