Add vignette about performance

[sebffischer]

As there are multiple ways to speed up training a neural network in torch, it would be useful to have a vignette that summarizes how to get these speed improvements.
This includes:

• jit compilation
• cuda
• ignite optimizers
• data loading:
  • pin_memory, non_blocking (https://pytorch.org/tutorials/intermediate/pinmem_nonblock.html?utm_source=chatgpt.com)
  • parallelization (and what it does not (yet) solve)