mruby-regexp: apply /i to the backreference comparison - #7046

A backreference ignored the /i flag: the RE_BACKREF case of bt_match()
compared the captured text with memcmp() unconditionally, so \1 stayed
case-sensitive even when the rest of the pattern was folded.

/(a)\1/i.match?("aA")           # CRuby: true, mruby: false
/(?<n>a)\k<n>/i.match?("aA")    # CRuby: true, mruby: false
/(a)(?i)\1/.match?("aA")        # CRuby: true, mruby: false
/(a)(?i:\1)/.match?("aA")       # CRuby: true, mruby: false

Fix

The case sensitivity is now decided at compile time, per instruction, the
same way every other ignorecase decision in this engine is taken.
compile_atom() records c->flags & RE_FLAG_IGNORECASE in the offset
field of the emitted RE_BACKREF (unused for this opcode), and bt_match()
selects an ASCII case-insensitive comparison when the bit is set.

Reading pat->flags from the matcher would have been a smaller change but
the wrong one. pat->flags is the whole-pattern option set, whereas (?i)
and (?i:...) are a compile-time mechanism on the compiler's own c->flags
that is never written back to the pattern. That form would fold
/(?-i:(a)\1)/i, which must not fold, and would leave /(a)(?i)\1/
unfolded, which must.

Reusing offset needs no change to re_inst. It is safe for this opcode:
insert_inst() and emit_atom_copy() rewrite offsets only for RE_JMP,
RE_SPLIT and RE_SPLITNG, and compute_fixed_len() rejects RE_BACKREF
through its default: arm.

Only bt_match() had to change. A pattern with a backreference always sets
has_backref, so mrb_re_exec() dispatches it to the backtracking engine;
the Pike VM has no RE_BACKREF case at all. literal_exec() is unreachable
here because is_literal requires !has_backref.

Case folding stops at ASCII, consistent with the existing /i handling for
literals and character classes. Extending /i to non-ASCII is a separate
limitation and is left untouched.

Tests

mrbgems/mruby-regexp/test/regexp.rb

• a new Regexp - backreference under /i assertion for \1
• /i cases added to Regexp - named backreference \k
• the inline forms /(a)(?i)\1/, /(a)(?i:\1)/ and the /(?-i:(a)\1)/i
  control added to Regexp - inline options (?i) / (?i:...)

All new assertions were run against CRuby 4.0.6 and produce identical
results. rake test passes: 1968 total, 0 KO.

Summary by CodeRabbit

• Bug Fixes

  • Fixed case-insensitive matching for numeric and named regular-expression backreferences.
  • Inline and scoped case-insensitivity options now apply correctly to backreferences.
  • Scoped case-sensitive settings correctly override an outer case-insensitive option.
  • Exact matching behavior remains unchanged when case-insensitivity is not enabled.

• Tests

  • Added regression coverage for toggled, scoped, numeric, and named backreference matching.