fix: replace vacuum delay with epoch grace period

[pathscale]

Summary

• replace vacuum's fixed 100 ms grace delay with a two-epoch quiescence protocol
• reserve reused free links with an RAII lease and keep the lease through index publication/unghosting in insert, CDC insert, reinsert, and CDC reinsert
• close reuse admission before advancing the epoch, then asynchronously wait for every prior-epoch lease to finish
• take the fragmentation snapshot after the grace period
• preserve the public EmptyLinkRegistry::pop_max signature

This closes the in-flight free-link reuse race that the old sleep attempted to cover. It does not claim a general reader epoch across index lookup and row deserialization; that broader read-vs-vacuum validation remains separate work.

Performance

Added simple_delete_insert_reuse, which measures the delete/insert churn path where the epoch atomics execute. Five alternating copied-binary pairs (30 samples, 0.5 s warm-up, 1 s measurement) produced paired changes of:

• +1.2%
• -1.3%
• +1.8%
• +0.0%
• -0.5%

Median change was approximately +0.0%, with no consistent regression, so this is not feature-gated. Vacuum also no longer pays the unconditional 100 ms delay.

Verification

• cargo test (117 unit; 324 integration passed, 5 existing ignores)
• cargo clippy --all-targets -- -D warnings
• cargo fmt --check
• regression: vacuum cannot finish its grace period while a prior-epoch reuse lease is alive
• regression: a reused-link lease remains alive until its caller completes publication

[pathscale]

Superseded by #187, which consolidated and merged the backend-selection and publication hardening work after the combined review and validation rounds.