[AVR32] Remove optimization of unaligned word loads

If we let unaligned word loads bypass the generic unaligned handling,
gcc may combine it with a swap.b instruction and turn it into a ldwsp
instruction, which does not work with unaligned addresses.

Revert the optimization to prevent the RNDIS driver from crashing.
Hopefully we'll figure something out later (it may be better to do the
optimization in gcc.)

Signed-off-by: Haavard Skinnemoen <hskinnemoen@atmel.com>

diff --cc include/asm-avr32/unaligned.h
Simple merge