Refactor/cleanup python internals

[RalfG]

Summary

This PR simplifies the Python/Rust boundary for annotation and target extraction, and keeps the correlate flow working with both raw and pre-annotated preloaded spectra. Final performance improved substantially after the Rust-side follow-up.

Changelog

Added

• Support for using Rust ms2pip_extract_targets(...) in the correlate pipeline.
• Support for carrying AnnotatedMS2Spectrum objects directly through the preloaded-spectrum path.

Changed

• Updated MS²PIP to consume ms2pip_compute_theoretical_mz(...) results directly from Rust without wrapping them again in np.array(...).
• Updated annotation calls so annotate_ms2_spectra(...) no longer receives seq_lens.
• Updated annotate_spectrum(...) to return AnnotatedMS2Spectrum instead of Python tuple-converted annotations.
• Updated MatchedSpectrum to store annotated_spectrum instead of peak_annotations.
• Updated _validate_and_extract_targets(...) to batch target extraction through a single Rust call.
• Updated correlate_single(...) to use Rust target extraction as well.
• Updated tests/test_spectrum_processing.py to validate Rust target extraction directly using AnnotatedMS2Spectrum / FragmentAnnotation.
• Bumped ms2rescore-rs dependency in pyproject.toml to >=0.5.0a3,<2.

Removed

• Python-side _annotations_to_tuples(...) from the active annotation/target-extraction flow.
• Python-side targets_from_annotations(...) from the active correlate flow.
• Redundant Python-side conversion of annotation results back into tuple structures when pre-annotated spectra are already available.

Fixed

• Restored / improved correlate performance after the initial API migration by following up on Rust-side bottlenecks (down from ~110 sec to ~75 sec for ~78k spectra).