Benchmark performance overhead of spy

[olevski]

We should have an idea of how much overhead using something like spy represents.

For this purpose I propose two tests:

1. Overhead for an optimized script/command @Panaetius mentions he has something in Rust for reading many files at once
2. Overhead for a python ML model. Here it would be nice to test out training an image classification model. We can either get a simple one from the academic team or find one online. For example the CIFAR-10 dataset is ~200MB and can be easily fed into some reference image recognition CNN.

The goal is to compare the increase in execution time from running with spy relative to running without it.

[Panaetius]

I used the imdb dataset from https://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz which is around 80mb but contains ~100k text files.
the rust code used for preprocessing it it just creates nice csv files out of it, I ran it once for train and once for test.
Only takes 0.1s to run in my machine, but then renku would take on the order of a day to check if the data is pulled from lfs (this was created to have that addressed) and to create metadata