Fix unbounded Queue growth when the queue never drains

[Evil0ctal]

Fixes #747

Queue.shift() only resets its backing array when the queue drains completely. The pool's open queue never drains while two or more connections sit idle in it: a connection is taken with shift() and pushed back when its query finishes, so every query leaves one dead slot behind, and remove() scanned all of them from index 0. Connection hand-off therefore gets slower with every query served since open last drained — the "gets slower until restart" in #747, and why it only shows up once a burst has opened more than one connection. The same pattern grows busy under load, the pending queries backlog, and each connection's sent queue.

The fix is in src/queue.js only:

• shift() compacts once the dead prefix is at least as long as the live entries (and longer than 64, so small queues aren't reallocated on every call). Each compaction copies no more elements than were shifted since the last one, so shift() stays O(1) amortized.
• remove() searches from the cursor. Slots below it are always undefined, so for any defined value the result is unchanged; only the scan gets shorter.

Keeping c.queue in sync in index.js would stop remove() from missing, but not the growth — sent is never remove()d, and the backlog grows the same way — so the callers are untouched.

• The first commit adds a test that fails on 3.4.9 with '3,,b,a,1,c,0,d,e,true' != '3,,b,a,1,c,0,d,e,10003'. The backing array has no public observable and the slowdown only becomes measurable past ~10^5 queries, so it's a Queue unit test rather than a database test: it passes an Array subclass as initial, which .slice() keeps via Symbol.species (as Result already relies on), so its push sees the backing array's size. It also checks a remove() miss, a non-head hit, and a hit at the cursor after a drain. No timing assertions; it runs in a few ms.
• The second commit is the fix.
• The full suite passes on ESM, CJS and Deno (265/265 each) against PostgreSQL 17 in a container set up like the CI workflow, including the second cluster on 5433. The new test also passes on Node 12, 14, 16, 18, 20, 21, 22, 23 and 24 (ESM and CJS). eslint src tests is clean.
• The patched Queue was differentially fuzzed against 3.4.9 over ~124M operations, biased toward the compaction boundary, with no divergence in any push, shift, remove or length result.

End to end, max: 10, 500k select 1 after a warm-up burst:

	open backing array	µs per query, first → last 50k
3.4.9	500,010	30 → 123
this PR	≤ 75	25 → 24

🤖 Generated with Claude Code

https://claude.ai/code/session_01BMNuLV7ceCupEDQmQFNADT