Keep binary bounds in nested add_variable_dicts and add_variable_matrix - #964

Fixes #963.

The bug only shows up with two or more index sets, which is probably why it works for some people and not others.

add_variable_dicts() and add_variable_matrix() turn cat=LpBinary into bounds [0, 1] and category LpInteger, but they do it by reassigning cat. With more than one index set they then recurse with that cat (now LpInteger) and the original lowBound/upBound (None), so the leaf call never sees LpBinary and creates integer variables with bounds (-inf, inf). A 1-D call never recurses and gets the right bounds, and add_variable_dict() doesn't recurse either.

On master (4.0.1):

prob = LpProblem("t", LpMaximize)
y = prob.add_variable_dicts("y", ([1, 2], [1, 2]), cat=LpBinary)
# y_1_1 ... y_2_2: cat=Integer, LB=-inf, UB=inf
prob.setObjective(lpSum(v for d in y.values() for v in d.values()))
prob.solve(HiGHS(msg=False))
# master:     status Undefined, no solution (the variables are unbounded)
# this PR:    status Optimal, objective 4.0

Change

Both functions now use a separate local for the converted category, so the recursive calls get the original cat and do the binary conversion at the leaf as before.

Tests

• New PuLPModelTest.test_binary_variables_with_several_indices builds a 2-D binary dict and matrix and checks every variable is Integer with bounds [0, 1]. On master it fails with AssertionError: -inf != 0; with the fix it passes.
• uv run --all-extras --no-extra mosek python -m unittest discover -s pulp/tests (Windows, Python 3.14): 1041 tests, OK (skipped=605).
• ruff check pulp and ruff format pulp --check pass. ty check pulp reports the same 2 warnings as on master (resource.getrusage in pulp/utilities.py, which doesn't exist on Windows).