Fix bug in NHWC Convolution with depthwise kernels

Changing input buffers caused out-of-bounds reads due to wrong
management of internal state and resulting miscomputation of input offset
passed to the microkernel.

Thanks @kimishpatel for investigation and the fix!

PiperOrigin-RevId: 358917686
diff --git a/src/operators/convolution-nhwc.c b/src/operators/convolution-nhwc.c
index de7aab0..3ae0077 100644
--- a/src/operators/convolution-nhwc.c
+++ b/src/operators/convolution-nhwc.c
@@ -1136,10 +1136,6 @@
       convolution_op->compute.range[1] = output_height;
       convolution_op->state = xnn_run_state_ready;
 
-      convolution_op->last_input = input;
-      convolution_op->last_input_height = input_height;
-      convolution_op->last_input_width = input_width;
-
       return xnn_status_success;
     }
     case xnn_ukernel_type_vmulcaddc: