feat: add lsp-ai support

[Robzz]

This PR adds support for the lsp-ai language server. Since the server is language-agnostic, I've set the defaults to attach to all filetypes, and always run in single-file mode. I do have a couple questions though:

1. As it is, this is not a full, functional configuration, as there are parts for which I don't think there are good default values, so I just want to be sure you're good with this as long as I document it, since from a quick look at the supported server list it looks like most (all ?) servers work out-of-the-box with the default config of an empty setup{} call.
   In particular, a user would need to select an inference backend, which could either be some API requiring an API key or some other user-managed inference server (no sane defaults for us to provide here), or have lsp-ai do inference itself with llama.cpp, which requires a GPU if you expect a decent latency for inline completions, as well as downloading a few GBs of model files from HuggingFace, which I don't think is a great default either.
2. The contributor docs state: In general, if a server requires custom client-side commands or off-spec handlers, then the server configuration should be added without those in lspconfig and receive a dedicated plugin such as nvim-jdtls, nvim-metals, etc.
   The server exposes a custom textDocument/generation endpoint. If I'm reading this right, this would fall under off-spec handlers, and does not belong here ? Asking since a quick search in the repo did surface a couple of configurations that do similar things, but this seems rare.

Marking as draft as I still need to write the server docs.