improve util.root_pattern()

[tmillr]

• Accept root patterns matching broken symlinks. If the dirent exists, that should be good enough.

• perf: don't check/resolve/follow symlinks when matching root patterns

• Remove pointless conditionals such as if match ~= nil then return match

• Remove pointless check for file existence (vim.fn.glob() already guarantees this)

• perf: avoid unnecessary table creations (e.g. used only to concatenate strings; string.format() or .. should be preferred for simple/static string concatenations)

• perf: avoid synchronous vim.loop.fs*() calls where possible (they tend to be slower than their builtin counterparts)

• Account for the fact that functions like string.match() and string.gsub() return multiple values (e.g. use return (s:gsub(...)) to return the modified string)

• Add types to some of the utils

[justinmk]

Most (or all) of util.lua is on the path to deprecation. However, improvements to it, especially simplifications, are still helpful because it makes it easier to understand what is missing from "upstream" Nvim stdlib.

[tmillr]

It's basically ready now, but going through the file one more time, I observed and decided to fix the bug in root_pattern(), only to go down the issue-ref rabbit-hole and realize that said bug is actually feature now.

The way root_pattern() was modified to prefer pattern order over directory depth feels very wrong to me, and I think the issue that caused it to get changed should have instead been:

1. Fixed locally (e.g. by removing certain patterns, or calling root_pattern() multiple times separately). There was never anything keeping people from calling it multiple times with certain (higher-precedence) patterns at a time if that was absolutely needed, and the algorithm would have been the same (as it is today).
2. Fixed by introducing another, separate, root pattern util which works differently (i.e. as it does today)

And it seems to have led to some ugly/hacky code in order to workaround the new (and IMO unintuitive) behavior, and it has caused at least a couple bugs as well (#3508).

But that change was over a year ago, and perhaps things were different then. The way I understand it now is that the root_dir callback drives/defines a workspace and its members/roots? Or in other words, it is the only way to achieve a multi-root/multi-member workspace (maybe some servers determine this internally on their own—e.g. rust workspace members are explicitly defined—but I know at least lua_ls doesn't and relies on the client). This is another reason why I think root_pattern() should've stayed the way it was. A .luarc.json in a nested subdir should reuse the server but also create a new/separate workspace member/context, and it should probably even do the same upon finding a nested .git. But there may be some nuance here that I am unaware of.

But anyway, even though it should definitely be changed back to how it was, I understand (now) that that could potentially cause some servers' configuration to break as they may have been changed in the meantime in order to adapt. So just let me know whether to change it, or leave it (hopefully to be addressed later).