Revise Windows implementation of splitdrive

[GunnarFarneback]

This PR makes three improvements to the Windows splitdrive implementation:

1. The matched regex is split into pieces and annotated.
2. Forward and backward slashes are considered equivalent. This fixes splitdrive doesn't respect UNC sharepoints when the path is constructed with forward slashes  #38492.
3. The patterns in the regex are reordered so that long UNC paths
   and long drive letters are once more recognized. This has been
   broken since splitdrive under windows #19695 (Julia 0.5).

Some things to discuss:

1. Do we want to support the long forms? They have been broken for a
   long time without causing much of a stir.
2. Why do we allow more than one drive letter? Should we do that?

[GunnarFarneback]

The second commit excludes reserved characters and requires drive letters to actually be one letter. I'm not entirely sure whether this is a good idea. On the plus side it makes the function more accurate, on the minus side it moves this closer to a breaking change. It also becomes stricter than Python's os.path.splitdrive but I'm not sure whether that's a plus or a minus. Feel free to convince me whether to keep or drop this commit.

[musm]

The second commit excludes reserved characters and requires drive letters to actually be one letter. I'm not entirely sure whether this is a good idea.

As far as I can tell, it's impossible for the drive letter to be longer than one letter as there can at most be 26 MS-DOS drives.

[DilumAluthge]

The failures in Windows CI look like they are related to the contents of this PR.

[DilumAluthge]

Error in testset path:
Test Failed at C:\buildbot\worker-tabularasa\tester_win32\build\share\julia\test\path.jl:68
  Expression: joinpath(S("foo"), S("bar:baz")) == "bar:baz"
   Evaluated: "foo\\bar:baz" == "bar:baz"
Error in testset path:
Test Failed at C:\buildbot\worker-tabularasa\tester_win32\build\share\julia\test\path.jl:68
  Expression: joinpath(S("foo"), S("bar:baz")) == "bar:baz"
   Evaluated: "foo\\bar:baz" == "bar:baz"

Error in testset path:
Test Failed at C:\buildbot\worker-tabularasa\tester_win64\build\share\julia\test\path.jl:68
  Expression: joinpath(S("foo"), S("bar:baz")) == "bar:baz"
   Evaluated: "foo\\bar:baz" == "bar:baz"
Error in testset path:
Test Failed at C:\buildbot\worker-tabularasa\tester_win64\build\share\julia\test\path.jl:68
  Expression: joinpath(S("foo"), S("bar:baz")) == "bar:baz"
   Evaluated: "foo\\bar:baz" == "bar:baz"

[GunnarFarneback]

Some archaeology indicates that this test was introduced in #22978, apparently with the intention that bar:baz should be interpreted as an absolute path. I'm fine with either backing out the requirement that a drive letter is a single letter or updating that test.

According to #20912 (comment), doing it like Python is preferred, or at least was five years ago. Can someone with a Windows Python investigate what

import os
os.path.join("foo", "bar:baz")

returns?

[vtjnash]

Since bar:baz is nonsense as a path construction, I would say chose any result that preserves the bar: appearing somewhere in the resulting path (effectively, making sure it is poisoned also). Since all options do that, you can change the test or the implementation–whichever you prefer.

[mzaffalon]

How does this PR deal with

julia> pwd()
"C:\\Users\\mzaffalon"

julia> joinpath(splitdrive(pwd()))
"\\Users\\mzaffalon"

ref: https://discourse.julialang.org/t/joinpath-fails-on-splitdrive-output/95328

[GunnarFarneback]

I expect splitdrive is already working according to specification for that path, and this PR has nothing to do with joinpath.

[mzaffalon]

On Julia v1.8.5 joinpath works fine for the following two cases

julia> joinpath("\\\\svfae01-solid01\\SED", "SE")
"\\\\svfae01-solid01\\SED\\SE"

julia> joinpath("C:\\", "Users\\mzaffalon")
"C:\\Users\\mzaffalon"

[musm]

bump?

[GunnarFarneback]

I'd encourage someone to pick this up. It's constantly below my attention threshold and I don't have a Windows machine myself.

[vtjnash]

Merged to master, while removing the second commit (which is breaking, and not really desirable to be that strict, as it means glob patterns won't be handled correctly anymore)

[vtjnash]

Is there somehow a bug in the MozillaCerts_jll that makes it hang similar to the REPL one that @Keno fixed recently?
https://buildkite.com/julialang/julia-master/builds/29602#018b8cb4-a4bd-4883-bc8a-d71361d13ef4