[mlir-vscode] Refactor server creation to be lazy

We currently proactively create language clients for every workspace folder,
and every language. This makes startup time more costly, and also emits errors
for missing language servers in contexts that the user currently isn't in. For example,
if a user opens a .mlir file we don't want to emit errors about .pdll files. We also don't
want to emit errors for missing servers in workspace folders that don't even utilize
MLIR.

This commit refactors client creation to lazy-load when a document that requires the
server is opened.

Differential Revision: https://reviews.llvm.org/D123184