Introduce 2-step classifier in alphaDIA scoring

[anna-charlotte]

WHO
Developers

WHAT
Replace the current neural network scoring in alphaDIA with a two-step classifier (LDA + NN).

WHY
Preselect precursors before NN training to speed up training and improve FDR sensitivity.

Acceptance Criteria
The solution must achieve at least the same FDR while reducing training time, but ideally should also improve FDR sensitivity by enabling better identification.

Additional information
The two-step classifier approach is inspired by DIA-NN by Demichev et al., as outlined in their Methods section.