Pool a matcher per thread

[headius]

Matching is a leaf operation, so we can pool a matcher per thread to avoid the cost of constructing and initializing the entire matcher for each string.

This patch does eliminate all object creation in typical usage, reusing all objects and expecting users to pull match data out of them before using the regex again. However in benchmarks of a regex-heavy StringScanner-based CSV library (jruby/jruby#7604) the overhead of fetching and clearing the pooled matcher appears to hurt performance more than it helps to remove the allocations.