[SPARC64]: Add -mtune=ultrasparc3 if possible.

This gives better heuristics for the cost of a multiply (fixed
5 cycles), rather than the 'ultrasparc' setting (variable, and
unpredictable if the second argument is non-constant).

Example code size savings:

   text	   data	    bss	    dec	    hex	filename
3823690	 304040	 448880	4576610	 45d562	vmlinux
3824521	 304040	 448880	4577441	 45d8a1	vmlinux.orig

Signed-off-by: David S. Miller <davem@davemloft.net>