Improve the comment watermark format

Watermarks are invisible strings inserted on CML comments[^1] so they can be differentiated from comments created by users. Our current watermark format is all but uniquely identifiable, and should probably be updated.

https://github.com/iterative/cml/blob/7bd925718ca2f1abb8d484fdd551a38683eb5cd3/src/cml.js#L181

Note that watermarks use a blank image instead of an HTML comment, because not all forges support the latter in their respective flavours of Markdown.

Proposal

Use the image description to store arbitrary data encoded as Base64 (e.g. useful for live metrics) and the image link to point to an invisible, tiny image in a static host of our choice.

![eyJrZXkiOiAidmFsdWUifQ==](https://cml.dev/9b0e332381649d87112c181dae1568640ea055b3100e5253d86410d276db8af9/pixel.gif) 

Precursors

• https://github.com/iterative/cml/pull/1026#discussion_r898745329

[^1]: E.g. created by running cml send-comment

[0x2b3bfa0]

We can also use Unicode zero-width watermarks: 🤔

• https://doi.org/10.1109/iThings-GreenCom-CPSCom-SmartData-Cybermatics50389.2020.00116
• https://github.com/offdev/zwsp-steg-js & https://github.com/enodari/zwsp-steg-py
• https://mayadevbe.me/posts/zw_steg

[casperdcl]

1. poor rendering: could this be fixed by using PNG instead of SVG?
2. poor URL: I’d be fine with replacing raw.githubusercontent.com/iterative/cml/master/assets -> static.iterative.ai/cml
3. guaranteeing no hidden tracking: the above (static.iterative.ai) is open sourced at https://github.com/iterative/static to prove no nefarious activity
4. base64: why? what is the use case? What metadata do we actually need to inject? If it’s just a comment ID then plaintext | urllib.parse.quote is enough?