1. The algorithm for applying edits was slightly incorrect. It needs to
preserve the original token list as the edits are applied instead of
mutating it as it iterates. From the spec:
Semantic token edits behave conceptually like text edits on
documents: if an edit description consists of n edits all n edits are
based on the same state Sm of the number array. They will move the
number array from state Sm to Sm+1.
2. Schedule the semantic token engine start() call in the
client._on_attach() function so that users who schedule_wrap() their
config.on_attach() functions (like nvim-lspconfig does) can still
disable semantic tokens by deleting the semanticTokensProvider from
their server capabilities.
* credit to @smolck and @theHamsta for their contributions in laying the
groundwork for this feature and for their work on some of the helper
utility functions and tests
