iommu: arm-smmu: optimize arch write sync delay

With ARM_SMMU_POWER_ALWAYS_ON enabled, clocks are always on. There arch
write sync is not required. Move arch write sync post checking for clock
refcount.
Arch write sync takes longer due to virtualized global register read. Use
non virtuazlied global register to reduce read latency.

Change-Id: Ic497c7ba75cc8ba2ed804f2c987e0c595487ce40
Signed-off-by: Prakash Gupta <guptap@codeaurora.org>