fix: release cloned mg_obj refcount after sv_magic

[Koan-Bot]

What

Fix a refcount leak on the cloned mg_obj in the magic cloning loop.

Why

sv_magic() takes its own SvREFCNT_inc on the obj parameter, but the
caller's reference returned by sv_clone() was never released. Each clone()
call on a tied variable leaked one refcount per magic entry — the cloned tie
object's DESTROY never fired.

This affects tied hashes, arrays, scalars, and any SV whose magic chain contains
a recursively-cloned mg_obj.

How

After sv_magic(clone, obj, ...), call SvREFCNT_dec(obj) when obj was
produced by sv_clone() (tracked via a per-iteration obj_cloned flag).
Shared mg_obj references (e.g. PERL_MAGIC_qr) are not affected — only
the cloned path needs the decrement.

Testing

• New t/28-tied-mgobj-leak.t with 5 tests covering tied hashes, arrays, and scalars
• DESTROY tracking confirms the leak existed (0 DESTROY calls before fix)
• Weakref-based detection verifies no surviving orphan after clone destruction
• Refcount assertion validates the expected count (magic + lexical = 2)
• Full test suite passes (29 programs, 367 subtests)

🤖 Generated with Claude Code

---

Quality Report

Changes: 3 files changed, 131 insertions(+)

Code scan: clean

Tests: passed (OK)

Branch hygiene: clean

Generated by Kōan post-mission quality pipeline