feat(broadcast): add an SPMC specialization

[onenewcode]

Summary

Closes #214.

• Add broadcast::spmc, a lossless single-producer broadcast family with a non-cloneable sender whose publish methods take &mut self.
• Keep the same public retention contract as broadcast::mpmc: bounded waits at capacity for the slowest active subscription, unbounded never waits, and a new subscription starts at the committed tail.
• Match the broadcast::mpmc public surface method for method, except the sender is not Clone. A receive drains accepted values before reporting disconnection. Unbounded value storage tracks the live window: discarded sends with no subscribers do not grow the log, and retained storage is released as the backlog drains.
• Keep subscriptions independent. Slot-log machinery stays private to asyncband/src/broadcast/spmc.

Design Notes

Public retention matches broadcast::mpmc: lossless fan-out, bounded wait-at-capacity, unbounded growth, join at the committed tail, and discard with no subscribers. try_recv drains accepted values before Disconnected. Unbounded discarded sends do not grow the log.

The sender is not Clone; publish takes &mut self. A blocked bounded send cannot subscribe; use resubscribe. The slot log stays family-private. Publication shares the waiter mutex with subscribe and parking; receivers drain published slots without that lock.

Benchmark comparison

One producer. broadcast::spmc senders are not cloned. mpmc 1P is broadcast::mpmc with one sender.

Bounded is the lossless wait-at-capacity path (async-broadcast overflow off). Tokio is omitted: it overwrites at capacity. Unbounded is the non-blocking batch path; Tokio and async-broadcast get capacity 4096 so the batch stays off Lagged. Native-thread concurrent waits use FutureExt::block_on. Figures are divan medians from one run on Apple M4 (cargo bench -p benchmarks --bench ecosystem -- broadcast::spmc).

Bounded (lossless wait-at-capacity)

Bench	async-broadcast	spmc	mpmc 1P
try_round_trip	24.6 ns	12.8 ns	32.4 ns
ready_round_trip	39.3 ns	16.2 ns	42.5 ns

Native threads (concurrent). Lower time is better.

Shape	async-broadcast	spmc	mpmc 1P
cap 1 / 1 recv	16.2 ms	9.25 ms	15.5 ms
cap 1 / 8 recv	123 ms	72.5 ms	76.1 ms
cap 1 / 32 recv	526 ms	216 ms	235 ms
cap 64 / 1 recv	491 µs	241 µs	537 µs
cap 64 / 8 recv	10.1 ms	21.8 ms	14.5 ms
cap 64 / 32 recv	37.8 ms	151 ms	171 ms

Tokio 4-worker (scheduled):

Shape	async-broadcast	spmc	mpmc 1P
cap 1 / 1 recv	1.28 ms	585 µs	908 µs
cap 1 / 8 recv	14.5 ms	3.59 ms	5.21 ms
cap 1 / 32 recv	67.1 ms	14.9 ms	29.6 ms
cap 64 / 1 recv	233 µs	113 µs	255 µs
cap 64 / 8 recv	2.14 ms	392 µs	2.13 ms
cap 64 / 32 recv	5.92 ms	1.02 ms	7.48 ms

Unbounded (non-blocking batch)

Bench	async-broadcast	tokio	spmc	mpmc 1P
try_round_trip	24.5 ns	17.4 ns	12.8 ns	32.5 ns
ready_round_trip	26.3 ns	21.5 ns	14.8 ns	36.1 ns

Fan-out (publish 4096, then every subscription drains):

Receivers	async-broadcast	tokio	spmc	mpmc 1P
1	127 µs	136 µs	87.4 µs	162 µs
2	235 µs	149 µs	120 µs	249 µs
4	476 µs	165 µs	142 µs	701 µs
8	907 µs	188 µs	179 µs	1.50 ms
32	3.32 ms	588 µs	436 µs	5.52 ms