[ARM] Context switch improvements

[ARM] Context switch improvements
 - We don't need to save the CPSR.
 - Rearrange thread_info members so we can pull the fields out of
   thread_info more efficiently.
 - Allocate a couple of extra words for CPU specific context saving
   (eg, for Xscale ACC registers)
 - Always leave 8 bytes free at the top of the kernel stack.  This
   prevents the stack becoming completely empty when do_exit() is
   called from an exiting nfsd() thread, and causing the wrong
   pointer to be returned from current_thread_info()