Blackfin: only flag L1 instruction for DMA memcpy

The performance difference from doing an 8 bit DMA memcpy vs an optimized
core memcpy can be pretty big when you add in the overhead of setting up the
MDMA registers, cache flushes, etc...  So only use dma_memcpy() when we
actually require it.

Signed-off-by: Mike Frysinger <vapier@gentoo.org>