Convert the expr codec from nightly coroutines to a manual iterator and push-struct

[MesTTo]

item_source, item_sink, and apply_e used nightly #[coroutine] generators. apply_e is
#[inline(never)] and received the sink as &mut Pin, so the coroutine resume stayed opaque
across that boundary (LLVM does not inline a resume across the Pin/poll edge), paying
state-machine overhead on a trivial per-item body.

This converts item_source to an ItemSource struct implementing Iterator, item_sink to an
ItemSink push-struct (push/finish; the old arity stack only existed to signal completion,
which apply_e never reads), and apply_e to take the sink directly, dropping the Coroutine/Pin
bound. The apply_e_clears_stacks_and_cycles_check macro keeps its by-name contract, so the
space.rs call sites are unchanged, and eval's quote loop becomes a plain for. Two
differentials pin the conversion: diff_item_sink_old_vs_new keeps a frozen copy of the old
coroutine sink and asserts byte-equal output and counts over a ground and adversarial corpus
(coreference, varrefs at the intro boundary, deep nesting, self-referential shapes), and
diff_apply_e_emit_vs_unify_oracle cross-checks the apply_e emit against the independent
Expr::unify substitution path.

Measured, min of 3 interleaved against the pre-change binary on one machine:

• bench process_calculus: 31,771,686 -> 28,729,938 us, 9.6% faster
• Nil's bfc chainer (the bench bfc program as a plain mork run MM2 file, proof size 19, 801
  steps, a 2.9M-atom space): 14,057 -> 11,279 ms, 19.8% faster
• bench finite_domain: 71,632 -> 67,313 us

Byte-identical along the way: every kernel/resources program at several step depths, the bfc
reproduction, and a 4096-hub triangle produce identical result spaces against the pre-change
binary. mork-expr tests 17/17.

Independent of #124; the two touch disjoint files.

[Adam-Vandervorst]

Hi! Appreciate your work here. The MORK kernel is closed to constant-time speed-ups until the asymptotic runtime has no known deficiencies. This prevents us from locking out much more impactful changes. Thanks for your understanding.

[MesTTo]

Closing per the maintainer's guidance that the MORK kernel is closed to constant-time speed-ups until the asymptotic runtime has no known deficiencies. The codec coroutine->iterator conversion is a constant-factor change, so parking it.