use map.split for semtokens range queries

[ailrst]

• cache the actual token map so we don't recompute it from the ast every time

Visibly smaller delay in Kate editor which implements semantic token range queries and is easy to set up the LSP for. I believe VSCode will also use range queries.

Still a big delay on the initial processing.