Optimization to semantic tokens extracting logic

[soulomoon]

another thing, though this doesn't have to be addressed in this MR and could be a follow up - now that we are traversing the entire document in order, do we still need to accumulate results into a M.Map Range HsSemanticTokenType? Couldn't we instead directly accumulate a [(Range, HsSemanticTokenType)]?

Originally posted by @wz1000 in #3958 (comment)