Blender Git Loki

Cycles X: Implement pass accessor for GPU

It replaces film_convert_half_float with more generic code path.

Currently only half4 destination is supported. No real stopper reason
to support float destination, is just a bit harder to verify at this
point as there will be no users of this code path.

Form quick benchmarks the new code seems to be 10-15% faster than the
film_convert_half_float when measuring overall display update time for
non-interop rendering. Partially because we no longer convert global
index to x/y and back to pixel index, partially due to smaller kernel.

Note that the timing is only taking display update into account, the
path tracing is not included into the speedup calculation (this means
that viewport is not that much faster, but it means we can update
viewpor4t more often now).

Differential Revision: https://developer.blender.org/D11481