Expand standalone bench suite to mirror comparative.rs and add rust-hpke memory footprint to README

[danieldia-dev]

This PR makes two improvements to the benchmark infrastructure:

1. benches/bench.rs is rewritten to functionally serve as a standalone mirror of the hpke_ng members in comparative.rs. Group names and benchmark IDs now match exactly (e.g. kem/x25519/hpke_ng/generate, x25519_chacha20_seal_sweep/hpke_ng/64). The previous bench.rs used a flat monolithic structure (bench_kem, bench_setup, bench_seal_open, bench_export) with ad-hoc group names that did not align with comparative.rs. The rewrite switches OsRng to a seeded ChaCha20Rng (matching the entropy source in comparative.rs), splits into per-KEM and per-ciphersuite functions, and fills the following coverage gaps: p256 derive_key_pair and decap (previously missing); k256 all four ops (previously absent entirely); PQ KEM groups (kem/xwing, kem/mlkem768, kem/mlkem1024) and their setup counterparts under #[cfg(feature = "pq")]; AES-128 seal and open sweeps; context_open (seal + open), matching the lockstep design in comparative.rs; export lengths expanded from [32, 64, 128] to EXPORT_LENGTHS = [16, 32, 64, 128, 256]; and a roundtrip_1k end-to-end benchmark. bench.rs also adds hpke_ng/auth_encap and hpke_ng/auth_decap inside kem/x25519, which are not present in comparative.rs. criterion_group! / criterion_main! are split into classical_benches and pq_benches with cfg gates so --features pq controls which targets compile. The new suite compiles and runs under:

RUSTFLAGS="-C target-cpu=native" cargo bench --bench bench
RUSTFLAGS="-C target-cpu=native" cargo bench --bench bench --features pq

2. The README performance section is updated to reflect the rust-hpke memory footprint data collected in the prior session. The memory table gains a rust-hpke column covering Context<_, _, ChaCha20Poly1305> (96 bytes), ExportOnly (184 bytes), Aes128Gcm (912 bytes), Aes256Gcm (1,168 bytes), and minimal release binary (430 KB), with a note explaining the AeadCtxS / AeadCtxR type equivalents, the KDF-dependent size growth, and why the ExportOnly row is larger in rust-hpke than in hpke-ng. Table column separators are also aligned throughout for readability.