epoch: Skip try_advance in collect() when the queue is empty

[raphaelroshan]

Addresses #1001 (and the overhead reported in #852).

Problem

Global::collect() runs on the pin() hot path (every PINNINGS_BETWEEN_COLLECT pins) and unconditionally calls try_advance() first. try_advance() is genuinely expensive: a SeqCst fence plus a full traversal of the intrusive linked list of every registered participant (cache-miss-heavy pointer chasing, as the TODO there notes).

Advancing the epoch is only useful to make queued SealedBags expirable. When the global garbage queue is empty there is nothing to expire and nothing for the subsequent try_pop_if loop to pop, so the entire try_advance cost is wasted. In read-mostly workloads where defer_destroy is rare (the scenario in #1001), this is paid on every periodic collect() and scales O(participants).

Change

Short-circuit collect() when the queue is observed empty, via a new Queue::is_empty(guard) (a single Acquire load of head.next, mirroring the existing test-only helper) — instead of the SeqCst fence + O(participants) traversal.

Why it's sound

Skipping the advance is monotone-conservative: it can only delay epoch advancement, never advance it early, so it can never cause premature reclamation. Liveness is preserved because any thread that pushed a bag is itself on the collect() cadence and will observe a non-empty queue and advance. Worst case is a bounded reclamation delay, in the same spirit as the existing IterError::Stalled early return that already declines to advance.

Verification

• cargo test -p crossbeam-epoch — lib (44) + doctests (51) green, including collector::tests::pin_holds_advance, incremental, and buffering.
• ci/crossbeam-epoch-loom.sh (loom, LOOM_MAX_PREEMPTIONS=2) — green.
• clippy (lib) clean; cargo fmt applied.

The win is deterministic by construction: in the empty-queue case collect() now performs one Acquire load instead of a SeqCst fence plus a traversal of all participants. Happy to add a benchmark or a counted-metric harness if useful.