fix: improve Memoize test pass rate (Storable::last_op_in_netorder, deep recursion warning)

[fglock]

Summary

Two independent fixes targeting failures seen in ./jcpan -t Memoize:

1. Storable::last_op_in_netorder — implements the flag reporting whether the most recent freeze/store used network byte order. Set by nfreeze()/nstore(), cleared by freeze()/store().
2. "Deep recursion on subroutine" warning — RuntimeCode now tracks per-sub recursion depth and warns (under the recursion warnings category) the first time depth exceeds 100, matching Perl's PERL_SUB_DEPTH_WARN behavior. Map/grep/eval blocks and builtins are exempt.

Memoize test results

Test	Before	After
t/correctness.t	StackOverflowError after test 16	17/17 pass
t/tie_storable.t	Died at test 6 (Undefined subroutine &Storable::last_op_in_netorder)	5/6 pass

The lone remaining tie_storable.t failure requires tied-hash DESTROY to fire when my %cache goes out of scope — a separate pre-existing limitation unrelated to these fixes.

Not included: threads->new / ->join stubs

An earlier version of this branch added new as an alias for create and a synchronous join that ran the thread body in the current thread. That was rejected as unsafe: whenever a caller relies on thread isolation, running the body synchronously silently produces incorrect results. Memoize's t/threadsafe.t test 8 is exactly such a case — the "thread" calls unmemoize('count_up'), and with a synchronous join that really unmemoizes count_up on the main thread, so the main-thread assertion fails. Leaving threads-using tests failing loudly is better than a convincing-but-wrong stub. t/tie.t / t/tie_db.t failures are caused by a broken user-local CPAN install of DB_File (which requires XS) and aren't addressed here either.

Test plan

• make passes (all unit tests)
• jcpan -t Memoize: unblocks correctness.t and tie_storable.t
• Deep recursion warning honors no warnings "recursion"
• Deep recursion warning does not fire for fib(20) or other non-pathological recursion

Generated with Devin