refactor(mpsc): replace the legacy queue backend

[tisonkun]

Parent: #206

Goal

Reconsider the existing bounded and unbounded MPSC implementation from first principles now that the direct source layout and benchmark baseline are established.

Direction

• Preserve the current public sender, receiver, and error contracts.
• Use a single-consumer notification path rather than a competing-consumer wake-all mechanism.
• Reuse cancellation-safe internal waiting primitives where they fit.
• Remove the legacy split and unsafe impl Sync if an owned queue core can provide the same API more directly.
• Keep bounded capacity and unbounded send semantics explicit.
• Keep MPSC-specific implementation details under asyncband/src/mpsc; move machinery to the private crate-level internal module only when another family actually shares it.

Acceptance criteria

• Cancellation, disconnection, draining, ordering, and backpressure behavior remain covered at observable boundaries.
• Benchmarks include at least 1 and 8 producers with one consumer for bounded and unbounded channels.
• The new implementation has no sustained order-of-magnitude regression against comparable channels and does not materially regress the existing bounded MPSC contention path.
• Any new unsafe code has a localized invariant and independent justification.
• The final implementation remains straightforward enough to reason about without relying on an unpublished generic channel framework.

The benchmark baseline in #208 and the source-layout follow-up in #234 are complete.