iommu: arm-smmu: Fix clock reference count error

When an atomic iommu domain attaches, an additional vote for both
clk_prepare, bus_bw, and regulator_enable must be held. The prior logic
only did this if the atomic domain was the first to attach to the iommu.
Fix this.

As a side effect, add reference counting for bus_bandwidth voting such
that a call to arm_smmu_enable_clock() followed by
arm_smmu_disable_clocks() will not always result in a bus bandwidth
vote of zero.

Change-Id: I7f88ea845a281c8c1def4f642e61262b53b60e1a
Signed-off-by: Patrick Daly <pdaly@codeaurora.org>