fix(univmon): reject layer seeds and row geometries that break hash independence

[zzylol]

UnivMon's error bounds (documented in #221) treat every hash a key meets as an independent random draw. Two construction-time gaps break that. This PR rejects both at construction. Valid shapes hash exactly as before, and no wire bytes change.

Fixed

• Overlapping row fields in CountL2HH. Row r reads its column from bits [r·b, (r+1)·b) of one 128-bit hash (b = ⌈log2 cols⌉) and its sign from bit 127 − r.
  • Once rows·(b+1) > 128, a deep row's column bits overlap the shallow rows' sign bits.
  • Once rows·b ≥ 128, the shift runs past the hash.
  • The ASAPv1 encoder and decoder already refused these shapes, but the constructor built them silently.
  • CountL2HH::with_dimensions_and_seed (and so init_univmon and the UnivMonPyramid constructor) now panics on them. So does the OctoSketch L2hhWorkerSketch, which used to check only rows·b ≤ 128 and so accepted the sign overlap (for example 11 × 2048).
  • One helper, check_l2hh_hash_bits, now serves the constructor, the worker and both wire doors.
• Layer seeds.
  • Layer i hashes at seed index i, and seed indices wrap modulo the 20-entry SEEDLIST.
  • Layer 19 therefore hashed with BOTTOM_LAYER_FINDER (19), the same seed whose bits decide which keys reach which layer. Layer 20 reused layer 0's seed.
  • MAX_LAYER_SIZE drops from 64 to BOTTOM_LAYER_FINDER (19). Layers 0..=18 use 19 distinct seeds, none of them the finder's. A const assertion pins BOTTOM_LAYER_FINDER < SEEDLIST.len().
  • UnivMonPyramid, and both UnivMon wire decoders through check_layer_size, share the cap.

Why reject instead of drawing more hash

• Columns: drawing a second 128-bit word per key would change which cells existing shapes use. Every accepted shape already fits: the default 5 × 2048 needs 60 bits, and 10 × 2048 or 9 × 4096 still fit.
• Seeds: the seed list is inlined in every ASAPv1 sketch's metadata (seed_list). Extending it would change the bytes of every sketch in the crate and every golden. Re-mapping layer seeds would break the wire contract that layer i hashes at seed index i.
• Capping at 19 layers changes nothing for any pyramid that was correct before. With L ≥ ⌈log2(F0_max/k)⌉ + 2, 19 layers cover F0 up to k·2^17, about 4M distinct keys at k = 32.

Behaviour change

• init_univmon and the CountL2HH constructor panic on the shapes above.
• The UnivMon decoders reject layer_size > 19, where they used to accept up to 64. Such a pyramid's layer 19 hashed at the finder's seed.
• MAX_LAYER_SIZE is public and its value changes.

Tests (all fail on main)

• countl2hh_rejects_column_bits_that_reach_the_sign_bits (11 × 2048 = 132 bits).
• init_univmon_rejects_column_bits_that_reach_the_sign_bits.
• a_univmon_worker_refuses_column_bits_that_reach_the_sign_bits.
• layer_seeds_are_distinct_and_never_the_layer_finder.
• init_univmon_rejects_a_layer_on_the_finder_seed.

Also updated:

• the_widest_accepted_univmon_geometry_still_addresses_distinct_rows goes from 11 × 2048 (which overlaps) to 10 × 2048.
• countl2hh_accepts_the_widest_disjoint_geometry is new.
• The two l2hh_wire encoder-rejection tests now build their out-of-budget sketch directly, because the constructor no longer can.

Golden files

None change. No golden uses more than 19 layers or a shape over the hash budget, and asapv1_golden passes unchanged.

Sequencing

This PR is independent of #221; they touch different functions and different sections of api_univmon.md and docs/tests.md, so they merge in either order.

#176 and #198

• refactor(univmon)!: ASAPv1 only — drop the proto envelope field, add goldens #176 changes the UnivMon proto envelope and adds 3-layer, 2 × 4 goldens. This PR touches neither, and both orders merge cleanly.
• The ASAPv1 spec already states the CountL2HH budget, (ceil_log2(cols)+1)·rows ≤ 128. It states no UnivMon layer bound, so the doc is left alone here so as not to collide with refactor(univmon)!: ASAPv1 only — drop the proto envelope field, add goldens #176's spec edits.
• If the Go decoder (ASAPv1 alignment: UnivMon #198) should match the new layer cap, it needs the same layer_size ≤ 19 check.

Verification

• cargo fmt --check.
• Clippy -D warnings and cargo test --locked for each of the four CI feature sets.
• RUSTDOCFLAGS="-D warnings" cargo doc.

[GordonYuanyc]

This makes perfect sense. LGTM!