fix: reject M < 2 at compile time (closes #32)

[Dani-giron]

Summary

Hnsw::random_level() computes the insertion level with -ln(uniform) / ln(M). With M=1, ln(1) = 0, so the division produces f64::INFINITY, which casts to usize::MAX instead of panicking. The first insert() then calls initialize(usize::MAX), whose loop tries to allocate that many layers, consuming all available memory until the process is killed. M=1 is a valid value at the type level, nothing rejected it before construction.

Changes

• src/controllers/hnsw.rs: Hnsw::new() now has a compile-time const { assert!(M >= 2, ...) }. M=1 (or M=0) no longer compiles, with an error pointing at the exact Hnsw::new() instantiation that violates it. Apotheosis::new() calls Hnsw::new() internally, so this covers both entry points with a single guard.

Confirmed with Ricardo that the correct boundary is M > 1 (equivalently M >= 2): M=0 does not diverge (insertion_level comes out 0, a different and unrelated correctness issue with zero-capacity neighbor arrays, not covered by this fix), only M=1 produces the ln(M)=0 divergence.

Test plan

Verified locally with the same commands CI runs: cargo clippy --all-targets --all-features -- -D warnings clean, cargo fmt --check clean, and all 44 tests passing with the default M.

Also verified both directions manually: a scratch binary instantiating Apotheosis<_, _, 1, 2, 10> failed to compile with the assert's message, pointing at the Hnsw::new() call site; a normal build with default generics (M=16) compiled and ran unaffected.