soc_core: optimize mem_decoder

soc_core: optimize mem_decoder

Non-optimized version was tested on 7-series and was additional resource usage
was not noticeable. This does not seems to be the case on iCE40 (see #220), so
hand optimize it. On 256MB aligned addresses, it should be equivalent to the
old decoder used by previously in LiteX.

The only requirement is that to have address aligned on size, which was already
the case. An assertion will trigger it this condition is not respected.