Improve PCRE2 match performance for JIT and interpreted

[straight-shoota]

This patch contains a number of individual steps that improve performance of PCRE2 matching greatly.

• Global MatchContext that is allocated only once and shared between all match executions. JIT stack is assigned via a callback which returns the appropriate thread-local stack. The callback won't be called if JIT is not used. This improve JIT performance.
• Cache MatchData per instance and thread. This avoids re-allocating the backtracking stack which improve interpreted performance
• Now that the pointer to MatchData does not leak outside the local scope, there's no need to allocate it with the GC. libpcre2 can manage its memory on its own and the MatchData instances are handled in the bindings. This reduces GC pressure.

Using the benchmark program from #13144 (comment), I get these results:

$ bin/crystal run .test/bm-pcre2.cr --release             # (master)
starts_with? 284.80k (  3.51µs) (±18.92%)  20.1kB/op   7.43× slower
    matches?   2.12M (472.51ns) (±114.77%)    128B/op        fastest
$ bin/crystal run .test/bm-pcre2.cr --release             # (performance/pcre2-match_context)
starts_with?  18.26M ( 54.77ns) (±18.56%)  0.0B/op   1.52× slower
    matches?  27.67M ( 36.14ns) (±16.79%)  0.0B/op        fastest
$ bin/crystal run .test/bm-pcre2.cr --release -Duse_pcre1 # (master)
starts_with?  19.73M ( 50.67ns) (±14.41%)  16.0B/op   2.22× slower
    matches?  43.73M ( 22.87ns) (±17.42%)   0.0B/op        fastest

This shows a great improvement in match performance. The PCRE1 implementation is still significantly more performant in JIT mode (matches?).
A factor for this could be that the PCRE1 bindings are not thread safe. I'll leave investigation into this as a follow-up and consider the main regression as resolved.

Resolves #13144