msm: kgsl: Fix preemption perfcounter block math

Fix the calculation for the address of the memory to save and restore
preempted performance counters - make sure there is enough room for
the entire block for each ringbuffer.  In the process, optimize the
preemption initalization sequence - move the IOMMU check to the A5XX
specific code, and only set the preemption enabled bit after everything
successfully initializes.

Change-Id: Ic0dedbad398786f07da87430ea1f0c44e09a77ea
Signed-off-by: Jordan Crouse <jcrouse@codeaurora.org>