40 transfer hpc

[philswatton]

This PR resolves #40. It does this by implementing the following:

• attack script generation now generates slurm instead of bash scripts
• adversarial image generation now uses GPU if available
  • one exception worth noting: dependency chains means that MPS can't be used for the BoundaryAttack. There's an exception in the code for this, but this shouldn't be a problem on the HPC
• measuring transfer success metrics now uses GPU if available
• added a utility function for determining which GPU if any to use for adversarial image generation (as this uses pytorch and accelerator="auto" is a lightning feature)
• README updated to reflect slurm use
• updated values of epsilon to optimise over
• matplotlib added for plotting of adversarial images. not used in source code or scripts, but was useful in designing the code and we'll likely need it later when writing the report
• numpy version restricted to stop one of foolbox's dependencies breaking
• some smaller fixes to previous code, removing superfluous lines

[lannelin]

Looks good! Some minor comments.

haven't verified runs on HPC, let's pick this up next week.

[lannelin]

lots of vals to try, what's the rough expected cost implication of each additional val? (for each attack)

[philswatton]

So in a test run on the HPC with the original five values (across 1000 images, 2x2=4 attacks) the script took 2 minutes 35 to run

Since once of these is 0, I'd assume about 40 seconds extra per (non-0) epsilon (while keeping the number of attack images and attacks constant)

[lannelin]

nice work!

[joannacknight]

I've run the tests locally, and they all passed. I can't see any issues in the code.