msm: kgsl: Improve GMU firmware loading time

Doing a regwrite of one dword with barrier over a loop is very
inefficient for block writes to an I/O mapped memory. Use
memcpy_toio() to improve this since it is better for block copy
to io memory.

Change-Id: I21aa6145caeeb777f27b90fe091c50af49948722
Signed-off-by: Akhil P Oommen <akhilpo@codeaurora.org>
Signed-off-by: Carter Cooper <ccooper@codeaurora.org>