The IOVA API uses a memory cache to allocate IOVA nodes from. To make
sure that this cache is available, obtain a reference to it and release
the reference when the cache is no longer needed.

On 64-bit ARM this is hidden by the fact that the DMA mapping API gets
that reference and never releases it. On 32-bit ARM, this is papered
over by the Tegra DRM driver (the sole user of the host1x API requiring
the cache) acquiring a reference to the IOVA cache for its own purposes.
However, there may be additional users of this API in the future, so fix
this upfront to avoid surprises.

Fixes: 404bfb78

 ("gpu: host1x: Add IOMMU support")
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>