arm: dma_alloc_coherent: malloc() -> memalign()

Even though the MMU/D-cache is off, some DMA engines still
expect strict address alignment.

For example, the incoming Faraday FTMAC110 & FTGMAC100 ethernet
controllers expect the tx/rx descriptors should always be aligned
to 16-bytes boundary.

Signed-off-by: Kuo-Jung Su <dantesu@faraday-tech.com>
CC: Albert ARIBAUD <albert.u.boot@aribaud.net>