Update: improve performance simplexDual

[AlbertoCuadra]

This change simplifies simplexDual by replacing the previous enlarged LP formulation with a shifted-variable formulation. Previously, the max-min problem was expressed by introducing explicit inequality constraints,

$$ x_j - t \ge 0, $$

which required constructing an additional inequality matrix using eye(n) and solving the expanded LP.

The new formulation writes

$$ x = z + t, $$

where $z \ge 0$ and $t \ge 0$ are handled directly by the simplex formulation. Substituting into the equality constraint gives

$$ A z + \left(\sum_j A_j\right)t = b, $$

with objective

$$ \max t, $$

implemented as

$$ \min -t. $$

This reduces the LP size, avoids constructing the explicit inequality matrix, and solves the problem directly using simplex.

The output contract remains unchanged:

$$ x_{\min} = t $$

and

$$ x = z + x_{\min}. $$

Additional cleanup includes normalizing b as a column vector and clamping the recovered x_min to zero to avoid small negative numerical artifacts.