Remove the NumBits and NumVals traits

[kampersanda]

Summary

• Remove bit_vectors::NumBits and int_vectors::NumVals, along with their impls and prelude re-exports.
• Add inherent num_ones() and num_zeros() to BitVector, Rank9Sel, DArray, and SArray. Lengths are available through the existing inherent len() / is_empty().
• WaveletMatrix now stores the number of zeros in each layer (zeros) and the sequence length (len), so its layer bound becomes Access + Rank + Select.
• Update module docs, internal call sites, and the bench crate accordingly.

Why

Both traits only reported size statistics and were an over-abstraction, in the same spirit as #129:

• NumVals had no generic users at all. Every impl just forwarded to the inherent len().
• NumBits::num_bits also just forwarded to len(). The trait name described only one of its three methods.
• The cost of num_ones was not part of the contract: O(1) for Rank9Sel, DArray, and SArray, but O(u) for BitVector.
• The only generic user was WaveletMatrix, which called num_zeros() on every layer in its query loops. These values are known at construction time, so storing them is the textbook layout (Z_l in Claude and Navarro's paper). It also avoids the O(u) scan per query when BitVector is used as a layer.

We also considered keeping a minimal length-only trait (Len) for future generic structures such as LOUDS or FM-index, but decided against it: such structures can store the values they need in the same way as WaveletMatrix.

Notes for reviewers

• Breaking change: public traits are removed. Code that names NumBits / NumVals in imports or bounds, or calls num_bits() / num_vals(), must switch to len(). Calls to num_ones() / num_zeros() on concrete types keep working since they are now inherent methods. This should go into 0.10.0.
• Serialization format change for WaveletMatrix: zeros and len are now serialized after layers and max_value, so data serialized by earlier versions cannot be read.
  • Deserialization returns InvalidState if zeros.len() differs from layers.len(), since zipping them would otherwise silently return wrong values (test_deserialize_inconsistent_zeros).
  • zeros could be recomputed at load time as rank0(len) if only len were stored, but that would require widening the Serializable bound to B: Serializable + Rank, cost O(u) per layer for BitVector layers, and panic at load time for DArray layers without enable_rank(). Storing zeros keeps the bound as B: Serializable. The overhead is 8 bytes per layer plus 8 bytes.
• EliasFano::from_bits now uses bv.is_empty() instead of comparing the length to zero (clippy len_zero).

Test plan

• cargo test --workspace
• cargo test --no-default-features
• cargo clippy --workspace --all-targets -- -D warnings
• cargo clippy --no-default-features -- -D warnings
• RUSTDOCFLAGS="-D warnings" cargo doc --no-deps
• cargo fmt --all -- --check
• cargo +1.62.0 check --lib (MSRV)

🤖 Generated with Claude Code