Use git command to fetch repositories instead of libgit2 for robust SSH support

[zanieb]

Closes #1775
Closes #1452
Closes #1514
Follows #1717

libgit2 does not support host names with extra identifiers during SSH lookup (e.g. github.com-some_identifier) so we use the git command instead for fetching. This is required for pip parity.

See the Cargo documentation for more details on using the git CLI instead of libgit2. We may want to try to use libgit2 first in the future, as it is more performant (#1786).

We now support authentication with:

git+ssh://git@<hostname>/...
git+ssh://git@<hostname>-<identifier>/...

Tested with a deploy key e.g.

cargo run -- \
    pip install uv-private-pypackage@git+ssh://git@github.com-test-uv-private-pypackage/astral-test/uv-private-pypackage.git \
    --reinstall --no-cache -v

and

cargo run -- \
    pip install uv-private-pypackage@git+ssh://git@github.com/astral-test/uv-private-pypackage.git \
    --reinstall --no-cache -v     

with a ssh config like

Host github.com
        Hostname github.com
        IdentityFile=/Users/mz/.ssh/id_ed25519

Host github.com-test-uv-private-pypackage
        Hostname github.com
        IdentityFile=/Users/mz/.ssh/id_ed25519

It seems quite hard to add test coverage for this to the test suite, as we'd need to add the SSH key and I don't know how to isolate that from affecting other developer's machines.

[MichaReiser]

Do we have any benchmarks that use git dependencies? Does it change the performance in significant ways?

[zanieb]

Do we have any benchmarks that use git dependencies? Does it change the performance in significant ways?

I don't think we do. Honestly I'm not sure it makes sense to take the time to measure right now because we need to support authentication. I think benchmarking should be done when considering #1786 instead.