Add CustomEvaluator implementation

[rlvandenbroek]

Add CustomEvaluator and Basic Custom Evaluator Tutorial

Summary

This PR introduces a flexible CustomEvaluator interface and corresponding CustomEvaluatorConfig, enabling users to supply arbitrary Python callables as scoring functions within the Thompson Sampling (TS) framework.

In addition, this PR adds a minimal custom_evaluator_tutorial Merimo notebook demonstrating:

1. Standalone usage of CustomEvaluator
2. Usage of CustomEvaluatorConfig
3. Integration into a full Thrombin Thompson Sampling workflow

The tutorial uses a simple example scoring function (heavy_atom_count) to provide a transparent, easily verifiable demonstration.

---

TODO before merging

• Integrate CustomEvaluator into test suite
• Integrate into production Thompson Sampling Merimo notebook (thompson_sampling_tutorial.py)

---

Motivation

Previously, evaluators were primarily lookup-based or hard-coded evaluators. This limited flexibility for:

• On-the-fly property evaluation
• Rapid prototyping of scoring functions
• Algorithmic or model-based scoring

CustomEvaluator enables direct injection of callable scoring logic while preserving compatibility with the existing ThompsonSamplingConfig pipeline.

---

Implementation Details

1. CustomEvaluator

• Accepts a scoring_function: Callable
• Evaluates RDKit mol objects directly

2. CustomEvaluatorConfig

• Mirrors existing config patterns (e.g., LookupEvaluatorConfig)
• Stores the scoring function
• Cleanly integrates with ThompsonSamplingConfig

3. Tutorial Notebook

notebooks/custom_evaluator_tutorial.py

Demonstrates:

• Small 5-SMILES sanity check
• Config-based evaluator instantiation
• Minimal TS run using full reagent dataset

---

Backward Compatibility

• No breaking changes.
• Existing lookup-based evaluators remain unaffected.
• TS API remains unchanged.