perf(matrices): cache MatrixAccessor properties with cached_property

[MaykThewessen]

Summary

Convert the remaining @property methods on MatrixAccessor (c, Q, sol, dual) to functools.cached_property. After #630, vlabels, clabels, A, b, sense, lb, ub, vtypes are already materialised once in __init__. The four remaining properties still recompute on every access, which shows up in solver paths that read the same attribute twice on one MatrixAccessor instance.

Also fix a small adjacent issue in expressions.BaseExpression._map_solution: m.matrices.sol followed by m.matrices.vlabels quietly constructed two MatrixAccessor instances (each running _build_vars + _build_cons). Binding to a local M = m.matrices reuses one.

Why this is safe

Each MatrixAccessor is short-lived: Model.matrices is itself a @property that returns a fresh instance on every access (linopy/model.py:350-352). So the cache lives at most as long as the local variable a caller holds, and the accessor never sees a mutated model. No new clean_cached_properties plumbing is needed.

Where the gain shows up

Multiple solver builders read the same property twice within one M = model.matrices:

• solvers.Highs._build_solver_model — M.c at line 1278, M.Q at 1299; both can be re-read in user code that inspects the accessor after solve.
• solvers.Gurobi._build_solver_model — M.Q at 1586, used again at 1587 in the QP branch.
• solvers.Mosek._build_solver_model — M.Q at 2860, again at 2861.

Repeat reads of M.sol / M.dual are also common in post-solve introspection.

Benchmark

Model: 360,600 variables, 1,200 constraints, solved with HiGHS (Python 3.12, scipy 1.17, numpy 2.4):

	cold (first access)	warm (repeat access)
master	1.0 ms	1.1 ms (recomputed)
patch	1.2 ms	0.4 µs (~2500×)

Repro script:

import numpy as np, pandas as pd, xarray as xr, time
from linopy import Model

N = 600
m = Model()
idx = pd.RangeIndex(N); jdx = pd.RangeIndex(N)
x = m.add_variables(xr.DataArray(np.zeros((N, N)), coords=[idx, jdx]),
                    xr.DataArray(np.full((N, N), 100.0), coords=[idx, jdx]), name="x")
y = m.add_variables(0, 100, coords=[idx], name="y")
m.add_constraints(x.sum(dim="dim_1") + y >= 10)
m.add_constraints(x.sum(dim="dim_0") >= 1)
m.add_objective(x.sum() + 2 * y.sum())
m.solve(solver_name="highs", io_api="direct", log_to_console=False)

for attr in ["c", "sol"]:
    M = m.matrices
    t = time.perf_counter(); _ = getattr(M, attr); cold = time.perf_counter() - t
    t = time.perf_counter(); _ = getattr(M, attr); warm = time.perf_counter() - t
    print(f"M.{attr}: cold={cold*1e3:.2f} ms  warm={warm*1e6:.2f} us")

Tests

• pytest test/test_matrices.py test/test_objective.py test/test_solution_lookup.py test/test_model.py test/test_constraints.py test/test_linear_expression.py test/test_quadratic_expression.py → 399 passed
• pytest test/test_optimization.py -k "not modified_model and not netcdf" → 241 passed, 2 skipped

(Pre-existing master failures unrelated to this PR: test_modified_model[*] and test_model_to_netcdf_* — both fail on upstream/master with the same scipy/numpy stack.)

Files touched

• linopy/matrices.py (+4 @cached_property, +1 import)
• linopy/expressions.py (1 line: bind m.matrices to local before reading two attributes)