Get a 3- to 4-fold speedup for sub()/subn(), split() and findall() by

not calling self.search(); instead, call self.code.match() directly
and interpret the list of registers it returns directly.  This saves
the overhead of instantiating a MatchObject for each hit, basically
inlining search() as well as group().  When a MatchObject is still
needed, one is allocated and reused for the duration of the scan.