Faster freeze of dense constraints into CSRConstraint #1009

Describe the feature you'd like to see

Child of #972 (phase 2, PR 6). This also helps dense models with freeze=True.

In a sparse model, most constraints still start as dense expressions. Variable bounds and storage balances are examples. CSRConstraint.from_dense (linopy/constraints.py:1590-1660) converts them, and this conversion is the largest remaining cost on the build side.

import pandas as pd, linopy
linopy.options["semantics"] = "v1"
m = linopy.Model(sparse=True)
y = m.add_variables(0, 1, coords=[pd.RangeIndex(2000, name="i"), pd.RangeIndex(720, name="t")])
m.add_constraints(y <= 1)  # 1.44M one-term rows: about 90 ms

A direct conversion from the term rectangle to CSR (a mask plus cumsum, no sort) builds the same matrix in about 36 ms.

Where the time goes:

• _matrix_export_data (linopy/constraints.py:2147-2173) makes several reshape and fancy-index copies. Its peak memory is about 4 times the size of the result.
• sum_duplicates sorts each row, even when every row has only one term and is therefore already canonical.
• np.unique over the per-row sign array checks whether all signs are equal.

Proposed direction

• Ravel vars and coeffs once, keep vars != -1 and coeffs != 0, and build indptr with cumsum.
• Canonicalise (sort and sum duplicates) only when nterm > 1. CSRConstraint must still promise label-ordered rows (v1 canonical form).
• Check for a uniform sign with (sign == sign[0]).all().
• Zero policy stays the same: zeros are dropped at freeze.
• Tests: frozen against dense equality for duplicate variables, explicit zeros, masked rows and mixed signs, using the existing helpers in test/test_csr.py.