Fix MaxPool and Clip operator

• Added second output for MaxPool and fix backward.
• Fixed Clip backward in edge cases;
• Fixed wrong filler type (Conv instead of Conv2D);
• Fixed non-accumulating gradient for Add and Div operators;
• Parallelized Conv and FC backward.

Some constrains in OptArgs are relaxed (ASSERT transformed in Log.warn()) to support Sequential with operators with optional outputs, like MaxPool.