iommu: bcm2712-iommu: Map and unmap multiple pages in a single call

[njhollinghurst]

For efficiency, the map_pages() and unmap_pages() calls now pay attention to their "count" argument, at least in successful cases.

Fix a bugette where we omitted to set mapped to 0 in case of error. Minor style fix.

[njhollinghurst]

This is a leftover from #5833. The efficiency improvement is admittedly quite modest; the average count (when exercised by camera) seems to be about 2.

Note that count does not feature in the large-page/alignment flags: because pages might get freed in a different order to the order in which they were allocated. We aren't bothering to check e.g. that pages are aligned to their own size, because we trust the framework to get it right.

There is no change in the "pass-thru" default case (which perhaps should be an error case, as it's never used), which still returns one page at a time. The bug fix in the "off-the-top" error case is needed, I think.

[6by9]

Looks reasonable based on my knowledge of it (ie very limited).
In theory a patch should do one thing, but this will all get squashed before upstreaming, so I don't care.

[njhollinghurst]

Tested with camera/ISP; I'll do a bit more testing with other Pi 5 blocks then de-draft it.

[njhollinghurst]

I'm going to push a tidied-up version that ignores special treatment of the "pass-through" region.
IOMMU-aware DMA drivers aren't told it exists and should never use it -- they should respect the force_aperture flag.
Meanwhile non-IOMMU-aware ones won't use this API.