test(benchmarks): add basic, local benchmark suite

[deepyaman]

Is local ML preprocessing with DuckDB faster than ML preprocessing with scikit-learn? How does one-hot encoding with IbisML on Snowflake compare to snowflake.ml.modeling.preprocessing.OneHotEncoder? Can ML preprocessing on the database outperform ML preprocessing in Ray Data? Should we have been training our deep learning models in T-SQL all along?

Let's start by looking at benchmarks at various data volumes and numbers of preprocessing steps locally. The purpose of this is to mostly understand the workflows wherein IbisML can provide value; it is not, for instance, to say that people shouldn't use scikit-learn for a lot of local ML pipelines.