perf: trim vacant arena tail slots

[tisonkun]

Summary

• replace the length-dependent vacant-list terminator with a fixed sentinel so the logical slot tail can shrink safely
• remove occupied tail slots directly and collapse a contiguous vacant tail while retaining the vector allocation
• cover tail collapse and free-list preservation with focused arena tests

Performance

broadcast::mpmc::unbounded::drain_with_receivers was measured from the merged main baseline and this branch with 30 samples of 20,000 iterations. Representative medians:

Fan-out	Baseline	This PR	Change
peak 32, live 1	48.19 ns	42.24 ns	-12.3%
peak 32, live 4	90.09 ns	81.53 ns	-9.5%
peak 256, live 1	113.5 ns	40.86 ns	-64.0%
peak 256, live 32	545.9 ns	478.8 ns	-12.3%

The peak 256, live 1 case now matches peak 1, live 1 (40.86 ns versus 40.81 ns), removing the historical receiver-peak component from the reclaim scan. Interleaved baseline/branch checks of event and completion single-waiter cancellation remained approximately flat.

Validation

• cargo x check
• cargo x test
• cargo x lint

[tisonkun]

Closing this version because the tail-trimming optimization targets a narrow workload and adds more free-list complexity than the measured absolute savings justify. I will follow up with a smaller change focused only on making the existing free-list representation easier to understand.