ENH: Add DoubleQuadratic datafit for asymmetric loss (#272)

[Aishwarya0811]

Context of the PR

Summary

Implements DoubleQuadratic datafit for asymmetric quadratic loss as requested in issue #272.

Changes

• New DoubleQuadratic class in skglm/datafits/_double_quadratic.py
• Support for both dense and sparse matrices
• Integration with existing penalties and solvers
• Comprehensive test suite
• Updated package imports

Mathematical Details

Implements asymmetric loss function:
L(ε,α) = (1/2n) * Σᵢ weights(εᵢ) * εᵢ²
where weights(εᵢ) = 2α + 2(1-2α) * 1[εᵢ>0]

• α < 0.5: penalizes positive residuals (overestimation) more heavily
• α = 0.5: equivalent to standard Quadratic loss
• α > 0.5: penalizes negative residuals (underestimation) more heavily

Testing

• Unit tests for all methods
• Verification that α=0.5 matches Quadratic exactly
• Integration tests with L1 penalty and AndersonCD solver
• Asymmetric behavior validation

Contributions of the PR

• Adds new asymmetric loss datafit requested by the community
• Enables modeling scenarios where overestimation and underestimation have different costs
• Maintains full compatibility with existing skglm infrastructure
• Provides foundation for asymmetric regression applications

Checks before merging PR

• added documentation for any new feature
• added unit tests
• edited the [what's new](- |Feature| Add :class:DoubleQuadratic datafit for asymmetric loss functions)