Copying acism structure to stack may be expensive

[wjjensen]

For my use case, callgrind highlighted the copying of acism structure from heap to stack in run_match() as occupying about 50% of the time spent overall in run_match(). How much this matters on real hardware, other architectures, or different use cases may vary but it's something to consider.

I replaced it with direct references to the passed in acism struct in the reverse_match branch of my fork