For small number of products, a single collection should be fine.
Having said that, keeping the inventory on a joined collection will allow you to just update those fields in the inventory collection without touching the full document. Even those typesense will only reindex changed fields, the whole document in still read from disk during updates to check which fields change. With a join collection approach the collection we fetch from disk will be much smaller so more efficient.
Gotchas of using joins: there are some limitations (we have to implement these missing features), for e.g. you can't query on the joined collection's fields, or facet on them yet.