Add member_order (fifo/lifo) + demand-history cull

[seriyps]

• Add member_order => lifo | fifo pool config (lifo is the default and preserves
  all existing behaviour). fifo uses an OTP queue as the free-member structure,
  handing out the longest-idle worker first (round-robin rotation).

• Replace the per-worker max_age idle-time cull gate with pool-level demand
  history: a fixed-capacity queue ring buffer (head = current bucket) tracks peak
  concurrent in-use count over a rolling max_age window. At each cull tick the pool
  sizes to that peak (floored at init_count). This gives LIFO and FIFO identical
  shrinkage behaviour for the same config — FIFO workers never go idle under rotation,
  so the old per-worker check never fired for them.

• max_age now means "demand memory window" rather than "individual worker idle time".
  Practical behaviour for LIFO pools is essentially unchanged; FIFO pools now cull
  correctly. cull_interval default changed from {1, min} to {15, sec} so the
  defaults satisfy cull_interval ≤ max_age.

• Fixed-size pools (init_count == max_count) skip demand tracking entirely
  (demand_buf = undefined) — cull never fires for them anyway.

• Hot upgrade v5 → v6: demand_buf is reconstructed from member time timestamps.

Benchmark (OTP 27.3, vs master baseline)

LIFO path overhead is within noise (0–5%) for all existing benchmarks; the cull-heavy
benchmark improves by ~8%. FIFO adds ~10–15% vs LIFO for single take/return
(queue vs list-head), negligible for drain-all and concurrent patterns.

[seriyps]

this can be rewritten to always use queue:out and in the true branch of the case below to use the original Q, but it won't give performance benefit because we keep current bucket on the queue head with in_r. The only time we may get an O(n) operation is when we call out_r, but it happens after we rotate len(Q) number of buckets, so amortized O(1) with no edge cases.