Address caveats katalog update script

We need to address (some of) the current caveats in the catalog update script:

- [ ] move script from /hack to /tools/python
- [ ] add developer documentation
- [ ] clone the `katalog` repo into a temp directory instead of relying on the user to have done it prior to running the script
- [ ] not all of the YAML files in the `katalog` repo may want to be included in the catalog upload, so alert the user to all the changes instead of relying on `git diff`
- [ ] read the `catalog_upload.json` file prior to regenerating it to capture assets that live outside the `katalog` repo and re-add them later

_Originally posted by @ckadner in https://github.com/machine-learning-exchange/mlx/issues/200#issuecomment-926195710_