Allow models to force a new batch

This is useful for a few things:
 - Work around bugs, such as having 2 images in one batch
 - Keep the image in a single batch for fully connected attention
 - Improve performance by not evaluating embeddings multiple times

model/input/input.go

@@ -15,6 +15,12 @@ type Input struct {
 	// stored in Multimodal, used for caching and comparing
 	// equality.
 	MultimodalHash uint64
+
+	// BatchBreak forces a new batch to be started with this
+	// input. For example, this can be used to align images
+	// with batches. Note that batches may be divided in additional
+	// locations as well.
+	BatchBreak bool
 }
 
 // MultimodalIndex is a multimodal element (such as an image)

model/models/gemma3/model.go

@@ -112,8 +112,8 @@ func (m *Model) PostTokenize(ctx ml.Context, inputs []input.Input) ([]input.Inpu
 			result = append(result, inp)
 		} else {
 			imageInputs := []input.Input{
-				{Token: 108},    // "\n\n"
-				{Token: 255999}, // "<start_of_image>""
+				{Token: 108},                      // "\n\n"
+				{Token: 255999, BatchBreak: true}, // "<start_of_image>""
 			}
 			result = append(result, imageInputs...)