The dma_cache_sync routines is used to flush caches for memory returned
by dma_alloc_attrs with the DMA_ATTR_NON_CONSISTENT flag (or previously
from dma_alloc_noncoherent), but the requirements for it seems to be
frequently misunderstood.  dma_cache_sync is documented to be a no-op for
allocations that do not have the DMA_ATTR_NON_CONSISTENT flag set, and
yet a lot of architectures implement it in some way despite not
implementing DMA_ATTR_NON_CONSISTENT.

This series removes a few abuses of dma_cache_sync for non-DMA API
purposes, then changes all remaining architectures that do not implement
DMA_ATTR_NON_CONSISTENT to implement dma_cache_sync as a no-op, and
then adds the struct dma_map_ops indirection we use for all other
DMA mapping operations to dma_cache_sync as well, thus removing all but
two implementations of the function.