Treat NaN as absent in v1 arithmetic

[FBumann]

Summary

• NaN in multiplicative/divisive constants now masks the term (makes it absent) instead of raising ValueError
• NaN in additive/subtractive constants is treated as 0 (additive identity)
• NaN in constraint RHS still raises ValueError — silently skipping constraints is too dangerous
• Variable.to_linexpr(nan_coeff) now correctly produces absent slots in v1 (was inconsistently filling with 0)
• New mixed-coordinate-arithmetic.ipynb notebook documenting patterns for combining variables with different coordinate subsets (from PR #591 feedback)

Motivation

Addresses @brynpickering's feedback in #591: the fillna(0) ceremony before every multiplication with incomplete cost data was verbose and error-prone. With NaN-as-mask, the pre-align pattern becomes clean:

# Before (NaN raises): fillna required everywhere
cap_a_al, cap_b_al, cost_a_al, cost_b_al = linopy.align(
    cap_a, cap_b, cost_a.fillna(0), cost_b.fillna(0), join="outer"
)
combined = cap_a_al * cost_a_al + cap_b_al * cost_b_al

# After (NaN masks): no fillna needed
cap_a_al, cap_b_al, cost_a_al, cost_b_al = linopy.align(
    cap_a, cap_b, cost_a, cost_b, join="outer"
)
combined = cap_a_al * cost_a_al + cap_b_al * cost_b_al

NaN behavior summary (v1)

Operation	NaN behavior
expr + nan	NaN → 0 (additive identity)
expr * nan	NaN → absent (masks term)
expr / nan	NaN → absent (masks term)
expr <= nan	ValueError (use .sel()/mask=)

Test plan

• All 5333 existing tests pass
• Verify notebooks execute: arithmetic-convention, missing-data, mixed-coordinate-arithmetic
• Review algebraic property consistency (NaN masking preserves commutativity, distributivity)
• Check edge cases: scalar NaN, array NaN, Variable vs Expression paths

🤖 Generated with Claude Code