mlx: mixed-precision quant and capability detection improvements (#15409)

Improve the MLX model creation pipeline with several model-agnostic changes:

- Rewrite supportsVision to use vision_config instead of architecture name
- Add supportsAudio for audio encoder detection
- Add alignment checking (isAligned) for quantization group sizes
- Support per-projection mixed quantization in MoE expert packing
- Record per-tensor quant metadata in safetensors blobs
- Parse per-tensor quant metadata at model load time
- Validate quantize output is non-empty before storing
- Fix pin/unpin cleanup in expert group quantization
- Promote v_proj/k_proj/down_proj to INT8 for INT4 base quant
- Add MetalIsAvailable() utility
- Skip audio encoder tensors from quantization

x/create/client/create.go

@@ -191,6 +191,10 @@ func inferSafetensorsCapabilities(modelDir string) []string {
 		capabilities = append(capabilities, "vision")
 	}
 
+	if supportsAudio(modelDir) {
+		capabilities = append(capabilities, "audio")
+	}
+
 	if supportsThinking(modelDir) {
 		capabilities = append(capabilities, "thinking")
 	}
@@ -496,32 +500,38 @@ func supportsThinking(modelDir string) bool {
 	return false
 }
 
-// supportsVision checks if the model supports image input based on its architecture.
-// Qwen3.5 multimodal checkpoints are published as ConditionalGeneration architectures.
+// supportsVision checks if the model has a vision encoder by looking for
+// vision_config in config.json.
 func supportsVision(modelDir string) bool {
-	configPath := filepath.Join(modelDir, "config.json")
-	data, err := os.ReadFile(configPath)
+	data, err := os.ReadFile(filepath.Join(modelDir, "config.json"))
 	if err != nil {
 		return false
 	}
 
 	var cfg struct {
-		Architectures []string `json:"architectures"`
-		ModelType     string   `json:"model_type"`
+		VisionConfig *map[string]any `json:"vision_config"`
 	}
 	if err := json.Unmarshal(data, &cfg); err != nil {
 		return false
 	}
 
-	for _, arch := range cfg.Architectures {
-		archLower := strings.ToLower(arch)
-		if strings.Contains(archLower, "qwen3") && strings.Contains(archLower, "conditionalgeneration") {
-			return true
-		}
+	return cfg.VisionConfig != nil
+}
+
+func supportsAudio(modelDir string) bool {
+	data, err := os.ReadFile(filepath.Join(modelDir, "config.json"))
+	if err != nil {
+		return false
 	}
 
-	typeLower := strings.ToLower(cfg.ModelType)
-	return strings.Contains(typeLower, "qwen3") && strings.Contains(typeLower, "conditionalgeneration")
+	var cfg struct {
+		AudioConfig *map[string]any `json:"audio_config"`
+	}
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return false
+	}
+
+	return cfg.AudioConfig != nil
 }
 
 // getParserName returns the parser name for a model based on its architecture.