copy(): Reuse existing buffers whenever possible

[karlrupp]

At present, viennacl::copy() from host to device for non-proxy objects (e.g. vector, matrix) results in writing to a new buffer. We should reuse existing buffers if possible for two reasons:

• The reallocation and destruction of buffers takes additional time
• If the object only wraps a host buffer, the object after copy() will have it's own buffer rather than work with the user-provided buffer.

[karlrupp]

Fixed here for dense matrices:
38aaea2
The use of injected handles with compressed_matrix is sufficiently well supported, no need for a more complicated copy() by now.