Refer to mirrors by name, path, or url (#34891)

With this change we get the invariant that `mirror.fetch_url` and
`mirror.push_url` return valid URLs, even when the backing config
file is actually using (relative) paths with potentially `$spack` and
`$env` like variables.

Secondly it avoids expanding mirror path / URLs too early,
so if I say `spack mirror add name ./path`, it stays `./path` in my
config. When it's retrieved through MirrorCollection() we
exand it to say `file://<env dir>/path` if `./path` was set in an
environment scope.

Thirdly, the interface is simplified for the relevant buildcache
commands, so it's more like `git push`:

```
spack buildcache create [mirror] [specs...]
```

`mirror` is either a mirror name, a path, or a URL.

Resolving the relevant mirror goes as follows:
    
- If it contains either / or \ it is used as an anonymous mirror with
   path or url.
- Otherwise, it's interpreted as a named mirror, which must exist.

This helps to guard against typos, e.g. typing `my-mirror` when there
is no such named mirror now errors with:

```
$ spack -e . buildcache create my-mirror
==> Error: no mirror named "my-mirror". Did you mean ./my-mirror?
```

instead of creating a directory in the current working directory. I
think this is reasonable, as the alternative (requiring that a local dir
exists) feels a bit pendantic in the general case -- spack is happy to
create the build cache dir when needed, saving a `mkdir`.

The old (now deprecated) format will still be available in Spack 0.20,
but is scheduled to be removed in 0.21:

```
spack buildcache create (--directory | --mirror-url | --mirror-name) [specs...]
```

This PR also touches `tmp_scope` in tests, because it didn't really
work for me, since spack fixes the possible --scope values once and
for all across tests, so tests failed when run out of order.

0 files changed, 0 insertions, 0 deletions