[2/n] use a slot-based impl for ItemSet

[sunshowers]

The motivation is that a vec is significantly more efficient than a hash table
at storing compact and compact-ish integer keys, as the benchmark data below
shows.

We use a slot-based implementation, i.e. Vec<ItemSlot<T>>, where
ItemSlot<T> is an enum with Occupied(T) and Vacant { next: ItemIndex }
variants. The Vacant entries form an embedded free chain, so that
inserts reuse vacant slots without any secondary allocation.

We also change shrink_to_fit to compact the slot buffer and rewrite internal
indexes.

Delta vs. the HashMap-backed ItemSet from the previous commit
(criterion median percent change; −N% is faster, +N% is slower):

get (point lookup, 32 B records):
    size  id_hash_own  id_hash_brw   id_ord_own   id_ord_brw
       1       −25.3%       −24.4%     +0.8% ⁎      +10.0%
      10       −40.4%       −40.8%       −13.5%       −10.0%
     100       −49.5%       −50.9%       −44.0%       −45.9%
      1k       −57.5%       −56.3%       −68.6%       −63.5%
     10k       −48.9%       −69.7%       −34.1%       −37.6%
     50k       −59.6%       −59.5%       −56.6%       −50.3%
    100k       −46.5%       −40.4%       −65.2%       −67.5%
    500k       +13.3%       +16.4%       −44.8%       −46.0%
      1M        −9.9%       −12.0%        −7.5%       +14.0%

  ⁎ p > 0.05, no statistically significant change.

get_large (point lookup, 1 KiB records):
    size  id_hash_own  id_hash_brw   id_ord_own   id_ord_brw
      1k       −45.7%       −45.3%       −44.4%       −46.1%
     10k       −41.6%       −49.0%       −33.5%       −20.8%
    100k      +7.5% ⁎       −32.2%        −8.0%       +12.9%
      1M       −19.7%       −68.2%       −14.5%        −7.4%

bulk_insert:
  size   id_hash    id_ord
   100    −29.2%    −25.3%
   10k    −30.8%    −29.6%
  100k    −33.3%    −27.4%

bulk_insert_large (1 KiB payload):
  size   id_hash    id_ord
    1k    −37.1%    −24.6%
   10k    −47.5%    −37.4%
  100k    −26.1%    −28.3%
    1M    −52.3%    −52.7%

churn (remove + reinsert at steady state, 1000 ops/iter):
  size   id_hash    id_ord
   100    −22.1%    −16.1%
   10k    −32.7%    −23.1%
  100k    −38.0%    −25.3%

iter (full traversal, 32 B records):
  size   id_hash    id_ord
   100    −36.6%     −9.8%
   10k    −34.9%    −24.0%
  100k    −39.2%    −39.8%

iter_large (full traversal, 1 KiB records):
  size   id_hash    id_ord
    1k    −32.1%    −16.4%
   10k    −38.7%    −52.4%
  100k    −23.2%    −61.9%
    1M    −33.6%    −67.2%

shrink_to_fit (50% holes, then compact):
  size   id_hash    id_ord
   100    −13.3%    −16.6%
   10k     −8.2%    −12.4%
  100k    −18.0%     +5.4%

ref_mut/id_ord_map: +5.4%.

Overall, this is a pretty nice win throughout! There are a few anomalies at large sizes that are best explained by L1 cache overflow, but in general we're up to 3x as fast as before, and significantly more competitive with the std maps, including being faster at a few things like bulk insert and iteration. (As a bonus, we also get to drop the rustc-hash dependency.)

[sunshowers]

The unsafe code here was reviewed at 2026-05-01, by Dan and ixi in the Rust Study Group meeting. Some notes:

• change from slab to something else since this isn't quite a slab allocator (I chose "slot")
• call the start of the slot buffer start_ptr to be clear