fix(interpreter): defensive cleanup for my-hash/my-array on JIT->interpreter fallback

[fglock]

Summary

Fixes a ClassCastException thrown by the bytecode interpreter at sub/scope
exit when a my %h or my @a is declared on a control-flow path that may
be skipped (early return, last, goto, or short-circuit &&/||
guarding the declaration), inside a sub that fell back from the JIT to the
interpreter.

class org.perlonjava.runtime.runtimetypes.RuntimeScalar
  cannot be cast to class org.perlonjava.runtime.runtimetypes.RuntimeHash
class org.perlonjava.runtime.runtimetypes.RuntimeScalar
  cannot be cast to class org.perlonjava.runtime.runtimetypes.RuntimeArray

Root cause

SCOPE_EXIT_CLEANUP_HASH / SCOPE_EXIT_CLEANUP_ARRAY blindly cast
registers[reg] to RuntimeHash / RuntimeArray. The compiler reuses
register slots across statements, so a slot later assigned to my %h
can transiently hold a RuntimeScalar left over from a prior
CREATE_LIST etc. If control flow skips the MY_HASH / MY_ARRAY
initialisation, the cleanup runs anyway and the cast fails.

The scalar variant SCOPE_EXIT_CLEANUP already had the matching
defensive instanceof check; this PR mirrors it for the hash and
array variants.

Trigger

Real-world repro: use Moose; -> Sub::Exporter::Progressive::import
uses goto \&Exporter::import which forces the whole import sub onto
the interpreter-fallback path. The sub also has lexical hashes/arrays,
so every Moose-based test died at use Moose; with the cast above.

Minimal repro:

sub t {
    my %h = (a=>1);     # SCOPE_EXIT_CLEANUP_HASH emitted
    print "ok\n";
    return;             # exits before the goto
    my $f = sub {1};
    goto $f;            # forces JIT->interpreter fallback
}
t();

Before this PR: class RuntimeScalar cannot be cast to class RuntimeHash.
After this PR: prints ok and exits cleanly.

What's in the diff

• BytecodeInterpreter.java -- guard the two cleanup opcodes with
  instanceof and add an extensive multi-paragraph comment explaining
  the invariants, the trigger and a copy-pasteable minimal repro, so
  the check is not "cleaned up" by a future refactor.
• New src/test/resources/unit/interpreter_myhash_myarray_scope_exit.t
  with 6 sub-tests covering my-hash, my-array, mixed, the
  Sub::Exporter::Progressive-style for-loop pattern, a stress
  loop, and a short-circuit-skipped my %h. All verified to actually
  exercise interpreter fallback via JPERL_SHOW_FALLBACK=1.

Test plan

• New regression test interpreter_myhash_myarray_scope_exit.t
  passes (6/6).
• All existing pre-PR tests in the new file actually trigger the
  JIT->interpreter fallback path (confirmed with
  JPERL_SHOW_FALLBACK=1).
• make (full unit-test suite, both backends) is green.
• use Moose; (after extracting Moose-2.4000) now loads
  Class::MOP and its dependency tree without the cast failure
  (next blocker is the unrelated XSLoader::load in Moose.pm,
  tracked in dev/modules/moose_support.md Phase 3).

Generated with Devin