Keep finite-difference residual evaluations within bounds

[yurekami]

jacobian_fd chooses an inward finite-difference direction, but its relative step can still cross the opposite bound when the interval is narrow. For example, with bounds [0, 1e-10], the default perturbation from zero evaluates the residual at about 1.49e-8. This can make least_squares fail for residuals defined only inside the supplied bounds.

Clip the actual evaluation points to the box and use their actual displacement as the finite-difference denominator. Step inward explicitly at the upper bound, including when an adjacent-float interval's midpoint rounds to that endpoint. Clipping the points themselves also avoids a rounding excursion across zero that can remain after clipping only the displacement.

The regression tests cover lower/interior/upper points, mixed parameter scales, intervals crossing zero, adjacent floating-point bounds, exact linear Jacobians, and a bounded least-squares solve. Before the fix, the nine new helper subcases and the solver regression fail. After the fix, the full minimizer suite passes: 18 tests and 13 subtests.

Validation used the current checkout's Python module and tests on Windows/Python 3.12 with installed MuJoCo 3.12.0 native bindings; no native code was changed or rebuilt. Reproduce that configuration from the repository root with:

from pathlib import Path
import mujoco
import pytest

mujoco.__path__.insert(0, str(Path('python/mujoco').resolve()))
raise SystemExit(pytest.main(['python/mujoco/minimize_test.py', '-q']))

Pyink checks pass for the changed regions, isort passes with mujoco classified as a third-party extension package, and the applicable pre-commit checks pass.