vulkan: disable integer dot product on AMD proprietary driver without…

[IIIIIllllIIIIIlllll]

Fix: #192
On AMD GPUs without native dot4 hardware (GCN/RDNA1/RDNA2, e.g. RX 6800), the AMD proprietary Windows driver advertises integerDotProduct4x8BitPackedSignedAccelerated=true, so ggml-vulkan routes q8_0 matmuls through the integer-dot path (dotPacked4x8EXT). That path silently produces wrong results there — with the default q8_0 Irodori-TTS package the output WAV is saturated to a constant +32767 with no error or abnormal metric. RADV reports accelerated=false on the same hardware, and RDNA3+ has native dot4. Same signature reported in ollama#13108 (RX 6650 XT).

Fix

• Force integer_dot_product=false for the affected combination (AMD vendor + eAmdProprietary + GCN/RDNA1/RDNA2), mirroring the existing old_amd_windows quirk condition. Equivalent to the known workaround GGML_VK_DISABLE_INTEGER_DOT_PRODUCT=1, applied automatically; also covers MUL_MAT_ID and flash-attention MMQ paths. The device-info path is adjusted too so the int dot log line stays truthful. RDNA3+, RADV, NVIDIA, Intel are unaffected.

Verification

• Reporter confirmed on the affected RX 6800: device now reports int dot: 0, q8_0/Vulkan output is bit-identical to the env-var workaround (25,462 distinct samples, RMS 0.1291), f16 unchanged, no measurable perf cost, test-backend-ops still 13,641/13,641.
• ggml-vulkan target compiles cleanly (MSVC 19.44, Vulkan SDK 1.4.350).

AI usage: Kimi K3 helped me find the issue and performed a fix.

[0xShug0]

Thank you @IIIIIllllIIIIIlllll! Great work! I'm really glad to have your help on the project.