kvcache: create cache ctx per layer

each cache layer creates and maintains its own context instead of using
a large context for all layers

ml/backend.go

@@ -99,7 +99,7 @@ type Context interface {
 
 	Forward(...Tensor) Context
 	Compute(...Tensor)
-	MaxTensors() int
+	MaxGraphNodes() int
 	Close()
 }