Gpu remove redundant copies

[pthomadakis]

This PR partially resolves #112

• First, the GPU kernel is examined and all load/store operations on any of its inputs are detected. Thus, we can track which arguments read, write or read-write the data provided by the host. As a result, the host can decide which data really need to be transferred to/from the device.
• Second, we add a simple and conservative heuristic that tracks memrefs used in each GPU kernel launch operation and tries to remove redundant data transfers when the host does not access the data used/ produced by a previous kernel execution. (Assuming a single GPU)