feat(memory): use QAT embedding model as default

Switch default local embedding model from embeddinggemma-300M to
embeddinggemma-300m-qat (Quantization Aware Training variant).

QAT models are trained with quantization in mind, which typically
yields better quality embeddings compared to post-training quantization.

src/memory/embeddings.ts

@@ -43,7 +43,8 @@ export type EmbeddingProviderOptions = {
   };
 };
 
-const DEFAULT_LOCAL_MODEL = "hf:ggml-org/embeddinggemma-300M-GGUF/embeddinggemma-300M-Q8_0.gguf";
+const DEFAULT_LOCAL_MODEL =
+  "hf:ggml-org/embeddinggemma-300m-qat-q8_0-GGUF/embeddinggemma-300m-qat-Q8_0.gguf";
 
 function canAutoSelectLocal(options: EmbeddingProviderOptions): boolean {
   const modelPath = options.local?.modelPath?.trim();