perf(v1): skip mask/where/const for dense variables in to_linexpr

[FBumann]

TODO: your own intent/motivation here.

Note

The following content was generated by AI.

What

Variable.to_linexpr under v1 unconditionally built an absent mask, a
where(~absent) copy, and an explicit full-size all-zero const array —
even for a fully-dense variable with no absent slots. On the isolated
var * array (exact-match) path these are all no-ops but still allocate
full-size temporaries.

This is the source of the remaining CodSpeed memory regression after #804/#816:

Mode	Benchmark	BASE	HEAD	Δ
Memory	test_op[var_mul_array_match]	212.3 KB	306.1 KB	−30.6%

and the small-model test_build peaks (knapsack-100 +19%, milp-10 +17%,
sparse_network-10 +16%, sos-10 +13%). Large models were already at parity —
the all-zero const is pruned at matrix assembly (#816), so built models
are byte-identical; the regression only shows on the isolated expression object.

Fix

Gate the mask / where / const on has_absence = (labels == -1).any().
Dense variables now build the same lean coeffs+vars expression as legacy;
masked variables still carry NaN at absent slots for §6 propagation.

Measurements

Isolated var * array peak (memray, warmed):

n	legacy	v1 before	v1 after
5 000	123.9 KiB	133.7 KiB	78.1 KiB
20 000	490.1 KiB	529.2 KiB	312.5 KiB
200 000	4884.7 KiB	5275.3 KiB	3125.0 KiB

v1 now sits below legacy on this path.

Tests

Full suite passes under both semantics: 7813 passed, 605 skipped
(pytest test/). Dense/masked equivalence verified: masked coeffs
still [0, nan, 2, nan].

Follow-up (not in this PR)

The v1 branch still runs reindex_like(fill_value=nan) even when the
coefficient already matches labels exactly (first_mismatched_dim returned
None). Skipping that no-op copy — mirroring _align_constant's
needs_data_reindex=False path from #804 — would trim further.