Previously this was implemented using a global cache, cache
this per-CPU instead and bump up the number of context IDs to
match NR_CPUS.
Signed-off-by: NPaul Mundt <lethal@linux-sh.org>