Normalize root-counting chain entries to survive f32 at high degree

[naseem173]

What & why

Fixes #295. root_counting_chain built up each new Sturm-sequence entry from raw pseudo-division remainders with no rescaling, so the coefficient magnitudes could drift geometrically down the chain. For a degree-12 polynomial with two close pairs of roots (0.0001 apart), that drift ran an f32 chain entry to exact zero partway through, and count_real_roots returned 8 instead of 12 -- the same failure mode described in the issue.

The fix renormalizes each new chain entry to unit max-norm (divide every coefficient by the largest absolute coefficient) right after it is built. sign_changes is the only thing that reads the chain back out, and dividing by a positive scalar never changes a sign, so this is a no-op for correctness on every case that already worked -- it only prevents the coefficients from running out of range.

Added count_real_roots_survives_high_degree_f32/_f64, which build the degree-12 polynomial described above and assert the count is 12 at both f32 and f64. The f32 variant fails on main (finds 8) and passes with this change; the f64 variant passes either way, since f64 has enough range that this particular case doesn't run out of it

Checklist

• cargo test + cargo clippy --all-targets clean locally
• New public APIs have a doc example
• No unwrap/expect/panic on library paths (typed errors instead)