Work on https://github.com/neovim/neovim/issues/27004.

`exec_lua` makes code slighly harder to read, so it's beneficial to
remove it in cases where it's possible or convenient.

Not all `exec_lua` calls should be removed even if the test passes as it
changes the semantics of the test even if it happens to pass.

From https://github.com/neovim/neovim/pull/28155#discussion_r1548185779:

"Note for tests like this, which fundamentally are about conversion, you
end up changing what conversion you are testing. Even if the result
happens to be same (as they often are, as we like the rules to be
consistent if possible), you are now testing the RPC conversion rules
instead of the vim script to in-process lua conversion rules."

From https://github.com/neovim/neovim/pull/28155#discussion_r1548190152:

"A test like this specifies that the cursor is valid immediately and not
after a separate cycle of normal (or an other input-processing) mode."

Problem:
LSP basic_finish test modified in #27899 is flaky again after #28030.

Solution:
Run on_setup() immediately after setup using a before_init callback.

Sets `NVIM_APPNAME` for the lsp server instances and also for the
`exec_lua` environment to ensure local user config doesn't interfere
with the test cases.

My local `ftplugin/xml.lua` broke the LSP test cases about setting
`omnifunc` defaults.

This fixes "fake server" from leaking memory, which makes ASAN very
upset, except on current ASAN CI for some reason.

The spec indicates that the response may be `null`, but it doesn't
really say what a `null` response means. Since neovim raises an error if
the response is `null`, I figured that ignoring it would be the safest
bet.

Co-authored-by: Mathias Fussenegger <f.mathias@zignar.net>