feat(problems): gradients for batch 10b's functions, and gradients and constraint values through Shifted and Rotated

[tachsin]

Closes #416.

Analytic gradients for 14 of batch 10b's functions (FitnessFunction::provides, evaluate_with):

• sum of different powers, the noiseless quartic, Penalized 1 and 2;
• the high-conditioned elliptic, bent cigar, discus, different powers;
• Büche-Rastrigin, Weierstrass, HappyCat, HGBat, Schaffer F7 and the rotated hyper-ellipsoid.

How they behave at the awkward points:

• Kinks follow Ackley's convention: where a term has a kink on a set of measure zero, its gradient there is 0. The kinks of different powers, HappyCat, HGBat and Schaffer F7 lie at their minima, so the gradient there is exactly 0.
• Weierstrass computes its phase from x, using sin(2πbᵏ(x + 0.5)) = −sin(2πbᵏx) for odd bᵏ. That makes its gradient exactly 0 at the minimum, where the naive formula leaves about 0.05 from rounding.
• The value always comes from evaluate, so it is identical to the bit.

No gradient, each with its reason in the docs:

• Step and the non-continuous Rastrigin (flat steps);
• Katsuura (kinks 2⁻³³ apart);
• the noisy quartic, whose noise makes it jump between any two genomes. It gives the gradient only when built without noise.

Wrappers:

• Shifted passes the wrapped problem's extras through unchanged.
• Rotated returns Mᵀ∇f(y) and J·M for the constraint Jacobian, at y = c + M(x − c).
• provides() reflects the wrapped problem's, so a shifted or rotated G06 keeps its constraint values, for Bayesian optimization.
• The Python adapter behind the wrappers passed on only the fitness; it now passes the gradient and constraint values too.

Tests:

• gradient::check at 200 random points per function, within 5e-8 beyond the differences' own rounding.
• Büche-Rastrigin by Richardson extrapolation.
• Weierstrass through its partial sums, since rounding in its top terms' phases defeats finite differences on the full sum.
• The wrappers' gradients and Jacobians against central differences, with the value bits unchanged.
• L-BFGS-B with supplied gradients on a rotated, shifted elliptic, to 6e-22.
• Python: gx.Lbfgsb, and gx.Bo with a wrapped G24.