perf(matrices): cache MatrixAccessor properties with cached_property (#716)

Convert the remaining `@property` methods on `MatrixAccessor` (`c`, `Q`, `sol`,
`dual`) to `@cached_property`. After #630, vlabels/clabels/A/b/sense/lb/ub are
materialised once in `__init__`, but these four properties still recomputed on
every access. Each `MatrixAccessor` is short-lived (rebuilt on every
`model.matrices` call), so cache invalidation is moot: the instance never sees
a mutated model.

Also fix a latent double-build in `expressions.BaseExpression._map_solution`:
`m.matrices.sol` followed by `m.matrices.vlabels` constructed two accessors
(each running `_build_vars`+`_build_cons`). Bind once to `M`.

Benchmark on a 360k-var / 1.2k-constraint model (Python 3.12, scipy 1.17,
numpy 2.4):

           cold (first)   warm (repeat)
  master:    1.0 ms          1.1 ms      (recomputed)
  patch:     1.2 ms          0.4 us      (~2500x on warm)

Repeat access happens in `solvers.Highs._build_solver_model` (M.c / M.Q both
read at lines 1278/1299 and again in Gurobi/Mosek paths at 1586-1587, 2860-2861)
and anywhere a caller assigns `M = model.matrices` then reads multiple times.

Author: MaykThewessen

linopy/expressions.py

@@ -969,7 +969,8 @@ def _map_solution(self) -> DataArray:
         Replace variable labels by solution values.
         """
         m = self.model
-        sol = pd.Series(m.matrices.sol, m.matrices.vlabels)
+        M = m.matrices
+        sol = pd.Series(M.sol, M.vlabels)
         sol[-1] = np.nan
         idx = np.ravel(self.vars)
         values = np.asarray(sol[idx]).reshape(self.vars.shape)

linopy/matrices.py

@@ -7,6 +7,7 @@
 
 from __future__ import annotations
 
+from functools import cached_property
 from typing import TYPE_CHECKING, cast
 
 import numpy as np
@@ -96,7 +97,7 @@ def _build_cons(self) -> None:
             np.concatenate(sense_list) if sense_list else np.array([], dtype=object)
         )
 
-    @property
+    @cached_property
     def c(self) -> ndarray:
         """Objective coefficients aligned with vlabels."""
         m = self._parent
@@ -121,7 +122,7 @@ def c(self) -> ndarray:
         np.add.at(result, label_to_pos[var_labels[mask]], coeffs[mask])
         return result
 
-    @property
+    @cached_property
     def Q(self) -> scipy.sparse.csc_matrix | None:
         """Quadratic objective matrix, shape (n_active_vars, n_active_vars)."""
         m = self._parent
@@ -130,7 +131,7 @@ def Q(self) -> scipy.sparse.csc_matrix | None:
             return None
         return expr.to_matrix()[self.vlabels][:, self.vlabels]
 
-    @property
+    @cached_property
     def sol(self) -> ndarray:
         """Solution values aligned with vlabels."""
         if not self._parent.status == "ok":
@@ -146,7 +147,7 @@ def sol(self) -> ndarray:
             result[positions] = var.solution.values.ravel()[mask]
         return result
 
-    @property
+    @cached_property
     def dual(self) -> ndarray:
         """Dual values aligned with clabels."""
         if not self._parent.status == "ok":