Figure out what we're doing with ICU tokenisation and locales

[matrixbot]

This issue has been migrated from #15124.

---

The ICU tokenisation rules seem to vary on different platforms.

Is it the ICU version? The locale? (How does ICU even get a default locale? I had a quick spelunk in the source code and couldn't find it!)

We need to figure out:

• what we actually want from the ICU library
• how we get that
• how we get consistent results.

It feels like we want a 'universal' locale independent of the host's settings, so that Synapse works well with all languages. (This may be a pie in the sky goal!)
What's the best we can do?

This issue was originally dug up in #15079, but e.g. Patrick's machine generates another tokenisation yet again. I'm not satisfied with the current solution..