This fixes k-takata/Onigmo#120.

The commit k-takata/Onigmo@9c13de8d0684ebde97e3709d7693997c81ca374b was insufficient.

https://github.com/k-takata/Onigmo/commit/1de602ddff140d91419e3f86dd35c81d7bd2d8e7

Fixes k-takata/Onigmo#92.

This fix was ported from oniguruma:
https://github.com/kkos/oniguruma/commit/257082dac8c6019198b56324012f0bd1830ff4ba

https://github.com/k-takata/Onigmo/commit/b1a5445fbeba97b3e94a733c2ce11c033453af73

https://github.com/ruby/ruby/actions/runs/16983250181/job/48147338699

For most tests (except two), we don't need to change Encoding.default_internal
in assert_raise_with_message. We're trying to run the test suite across
multiple Ractors and modifying Encoding.default_internal can cause other
concurrently running tests to fail.

... to conform to UTS 18 as mentioned in https://bugs.ruby-lang.org/issues/19417#note-3

https://unicode.org/reports/tr18/#word states word should match join_control chars.

It currently does not:

```ruby
[*0x0..0xD799, *0xE000..0x10FFFF].map { |n| n.chr 'utf-8' } => all_chars
all_chars.grep(/\p{join_control}/) => jc
jc.count # => 2
jc.grep(/\p{word}/).count # => 0
```

Prevent double free for too big repetition quantifiers

The previous implementation calls `free(node)` twice (on parsing and compiling a
regexp) when it has an error, so it leads to a double-free issue. This
commit enforces `free(node)` once by introducing a temporal pointer to hold
parsing nodes.