Normalize the Sturm chain entries so high degrees survive f32 #295

What: root_counting_chain builds each entry as the negated remainder of the two before it,
with no rescaling, so the coefficients grow or shrink geometrically down the chain. By degree ten
the later entries overflow to infinity in f32 and underflow to zero in both precisions, and since
sign_changes skips exact zeros and reads signs off the rest, count_real_roots and
real_roots_in then miscount silently. Only the signs matter, so normalizing each entry to unit
max-norm as it is built costs one pass and removes the failure entirely.

Where: crates/multicalc/src/polynomial/roots.rs.

How to test: cargo test -p multicalc; assert a degree-12 polynomial with known real roots is
counted correctly at both f32 and f64, which fails today at the higher degrees.