drm/nouveau: avoid GPU page sizes > PAGE_SIZE for buffer objects in host memory

While the Tegra (GK20A, GM20B, GP10B) MMUs support large pages in host
memory, we're currently lacking IOMMU support for merging system pages
into large enough chunks to be mapped as such by the GPU.

The core VMM code actually supports automatically determining the best
page size to map with, which is intended for these situations, but for
various complicated reasons the DRM is currently forcing the page size
selection on a per-BO basis.

This should fix breakage reported on Tegra GPUs in the meantime, until
one or both of the above issues are resolved properly.

Reported-by: Mikko Perttunen <cyndis@kapsi.fi>
Fixes: 7dc6a446 ("drm/nouveau: improve selection of GPU page size")
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Tested-by: Thierry Reding <treding@nvidia.com>