feat: optimize metric mappings

[gcalcedo]

Resolves #50.

This PR includes considerable changes to the core functionality.

• Metric maps are now written to filesystem as a dataframe with every sampling combination, instead of writing the complete map. This results in a CSV size orders of magnitude smaller.
• Metrics now use **kwargs instead of explicit arguments. The motivation comes from having metrics require different inputs, but still needing a homogeneous invoking call.
• The complete BERTopic pipeline is removed from the metric sampling. Instead, only clusters are computed in this step. Similarly, dimensionality reduction is now performed once per run of the MetricMapper. Both of these changes result in a substantial performance improvement.
• A progress bar with tqdm is added for better visualization.