fix: handle both string and array merge formats in tokenizer

Gemma 3 tokenizer.json uses [["a","b"],...] format for merges
instead of the ["a b",...] format. Support both.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

pkg/mlx/tokenizer/tokenizer.go

@@ -31,7 +31,7 @@ type tokenizerJSON struct {
 	Model struct {
 		Type         string          `json:"type"`
 		Vocab        json.RawMessage `json:"vocab"`
-		Merges   []string        `json:"merges"`
+		Merges       json.RawMessage `json:"merges"`
 		ByteFallback bool           `json:"byte_fallback"`
 	} `json:"model"`
 	AddedTokens []struct {
@@ -69,13 +69,29 @@ func Load(path string) (*Tokenizer, error) {
 		t.invVocab[v] = k
 	}
 
-	// Parse merges
+	// Parse merges — supports both ["a b", ...] and [["a","b"], ...] formats
-	for rank, merge := range tj.Model.Merges {
+	if len(tj.Model.Merges) > 0 {
+		// Try array-of-strings first
+		var stringMerges []string
+		if err := json.Unmarshal(tj.Model.Merges, &stringMerges); err == nil {
+			for rank, merge := range stringMerges {
 				parts := strings.SplitN(merge, " ", 2)
 				if len(parts) == 2 {
 					t.merges = append(t.merges, mergePair{a: parts[0], b: parts[1], rank: rank})
 				}
 			}
+		} else {
+			// Try array-of-arrays: [["a","b"], ...]
+			var arrayMerges [][]string
+			if err := json.Unmarshal(tj.Model.Merges, &arrayMerges); err == nil {
+				for rank, pair := range arrayMerges {
+					if len(pair) == 2 {
+						t.merges = append(t.merges, mergePair{a: pair[0], b: pair[1], rank: rank})
+					}
+				}
+			}
+		}
+	}
 
 	// Parse special tokens
 	for _, tok := range tj.AddedTokens {