I've run across a situation where Eglot receives a documentation
string as part of a "completionItem/resolve".  The documentation is
being provided as a regular "string", not MarkupContent, yet it is
being rendered as markup.  Since the string contains characters which
are being interpreted as markdown (e.g., ":"), it causes the
documentation to be rendered incorrectly.

[jsonrpc] e[19:22:22.458]     <-- completionItem/resolve[21]
{"jsonrpc":"2.0","id":21,"result":{"label":"End_Of_Line","documentation":"End
of line sequence: lf | crlf","data":["Format","End_Of_Line"]}}

I believe this is because `eglot-completion-at-point` calls
`eglot--format-markup` for the documentation string provided in the
CompletionItem without first checking to see if it really should be
formatted (i.e., if it is a string it shouldn't be formatted).  While
`eglot--format-markup` will handle strings, it should only be passed
strings which are intended to be markup formated (specified as
MarkedString in the specification).

I believe this issue might affect more than just the documentation
string for CompletionItem, it's possible this also affects other
places which specify fields as "string | MarkupContent" (such as the
SignatureInformation documentation field).