feat: Target Conditional generation framework

[Ramlaoui]

Builds on @andaero and @smglsn12, and expands on the possible conditional metrics to write. This introduces a framework for both:

• Discrete target generation where the goal is to exactly target one value with an equality check
• Continuous target generation, with the possibility to select the top k structures amongst the list, an option to turn off the tolerance threshold (which uses directly the distance as a metric then) or to use it for a success rate. It also supports various distance metrics that can be expanded

[vict0rsch]

Although not critical, I would consider using arrays rather than lists in such operations.

values = np.array(values, dtype=float)
valid_values = values[~np.isnan(values)]
...
success_values = (valid_values == self.config.target_value).sum()
success_rate = success_values.mean()
... # etc.

[vict0rsch]

Same, working with arrays would be faster.

Even such a simple manipulation is 10x faster with arrays:

In [1]: import numpy as np

In [2]: data = np.random.randint(0, 100, (10000,)).astype(float)

In [3]: data[data<5] = float("nan")

In [4]: values = data.tolist()

In [5]: %timeit [v for v in values if not np.isnan(v)]
3.73 ms ± 67.9 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [6]: %timeit (w:= np.array(values))[~np.isnan(w)]
267 µs ± 1.33 µs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)