ray_cs fixes

bugs galore:
	* 0xf380 instead of htons(ETH_P_AARP), etc.  Works only on l-e.
	* back in 2.3.20 driver got readb() and friends instead of
direct dereferencing of iomem.  Somebody got too enthusiatic and replaced
	ntohs(p->mrx_overflow)
with
	ntohs(read(&p->mrx_overflow)
without noticing that (a) the sucker is 16bit and (b) that expression can't possibly
be portable anyway (hell, on l-e it's always less than 256, on b-e it's always a
multiple of 256).  Proper fix is
	swab16(readw(&p->mrx_overflow)
taking into account the conversion done by readw() itself.  That crap happened
in several places; the same fix applies.
	* untranslate() assumes little-endian almost everywhere, except for
the code checking for IPX/AARP packets; there we forgot ntohs(), so that part
only works on big-endian.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: John W. Linville <linville@tuxdriver.com>