Enhance SMILES decoding and inference capabilities

[T4ras123]

This pull request introduces a flexible and extensible decoding framework for molecular conformer generation, allowing support for multiple serialization schemes (cartesian, cartesian_binned, uniform, quantile) and configuration via CLI or config files. It refactors the decoding logic in both single and multi-conformer inference pipelines, introduces a new BinConfig dataclass for bin configuration, and expands test set and CLI options.

Key changes:

Decoding Framework & Logic

• Introduced a unified decode_conformer_by_serialization function in smiles_encoder_decoder.py to handle decoding for various serialization schemes, including new support for "uniform" and "quantile" binning using configuration files. (decode_conformer_by_serialization, BinConfig, decode_cartesian_with_config) [1] [2]
• Added normalize_serialization_tag to standardize and validate serialization mode input.
• Updated all inference and multi-conformer inference pipelines to use the new decoding interface, passing serialization tags and bin config paths as needed. [1] [2] [3] [4] [5]

Configuration and CLI Improvements

• Added CLI arguments and config support for specifying serialization scheme (serialization_tag) and paths to uniform/quantile bin configuration files, both for single and multi-conformer inference. [1] [2] [3] [4] [5]
• The inference pipeline now auto-detects and harmonizes binned and serialization_tag to ensure consistent decoding. [1] [2]

Test Set Handling

• Added support for the "revisited" test set and improved handling for various test set formats in inference. [1] [2]

Utility and Robustness

• Enhanced strip_smiles to remove [SERIALIZATION]...[/SERIALIZATION] tags for cleaner input.
• Added missing imports and improved type annotations for clarity and maintainability. [1] [2] [3]

These changes make the codebase more modular, easier to extend with new serialization schemes, and more user-friendly for both research and production settings.