bench: cross-thread fanout, including the case this crate loses

[userFRM]

The in-cache benchmark isolates protocol cost, but cannot measure a design whose builder owns its consumer threads — that shape has no single-threaded mode, so the most relevant broadcast comparison was missing entirely. This measures the deployed arrangement: one producer, N consumer threads on their own cores, 100k messages, every consumer accounting for all of them.

Total milliseconds, lower is better:

	semantics	N=1	N=2	N=4	N=8
photon	lossless	0.49	0.52	0.58	2.62
sequence-barrier ring	lossless	0.45	2.94	6.99	12.15
per-consumer queue	lossless	1.39	14.90	27.92	46.79
shared-ring broadcast	lossy	8.23	8.61	33.86	119.47

At one consumer the barrier ring wins. No contention for a barrier to lose to, and its handler drains a batch where this crate takes messages one at a time. That stays in the table. From two consumers onward the positions reverse, and photon holds close to flat through N=4.

Two things the numbers do not establish, published alongside them:

• photon's own N=8 jump (0.58 → 2.62) may be scheduling rather than architecture — nine threads on a sixteen-thread machine that was not idle. Needs a pinned rerun on a quiet box before anyone relies on it.
• the shared-ring row is lossy by construction: a receiver that falls behind is told what it missed rather than holding the producer back. It belongs against this crate's lossy channel(), not the lossless rows. Writing this surfaced it — an earlier version asserted lossless delivery for every case, and that design cannot provide it.

Reproduce with cargo bench --bench fanout_threaded.

Verified: 219 tests, clippy clean, fmt clean.