Currently, we perform a request to Gitlab Rails that proxies
the request to primary

However, it causes timeouts on big pushes and consumes large
amount of memory. We can perform an HTTP request directly
from Gitlab Shell instead and stream the response to the user

All the implementations of DNS discovery were done in this epic:
https://gitlab.com/groups/gitlab-org/-/epics/8971. Gitaly allows clients
to configure DNS discovery via dial option. This MR adds the exposed
dial options to client connection creation in Gitlab-shell.

Issue: https://gitlab.com/gitlab-org/gitaly/-/issues/4722
Changelog: added

This commit also excludes gitlab-shell from dependencies:

Gitaly specifies Gitlab Shell as a dependency as well in order
to use gitlabnet client to perform API endpoints to Gitlab Rails.
As a result, Gitlab Shell requires Gitaly -> Gitaly requires an
older version of Gitlab Shell -> that version requires an older
version of Gitlab Shell, etc. Let's use exclude to break the
chain earlier

While gitlab-shell currently has a major version of v14, the module path
it exposes is not using that major version like it is required by the Go
standard. This makes it impossible for dependents to import gitlab-shell
as a dependency without using a commit as version.

Fix this by changing the module path of gitlab-shell to instead be
`gitlab.com/gitlab-org/gitlab-shell/v14` and adjust all imports
accordingly.

Changelog: fixed

Before this change, the GitLab internal API could use a boolean
response field to indicate whether gitlab-shell should make
sidechannel connections go Gitaly. We now ignore that response field
and always use sidechannel connections.

When gitlab-sshd has been introduced we've started running our
own SSH server. In this case we're able to cache and reuse
Gitaly connections and Registry.

It helps to reduce memory usage.

If the GitLab API returns an allowed response with use_sidechannel set
to true, gitlab-shell will establish a sidechannel connection and use
SSHUploadPackWithSidechannel instead of SSHUploadPack. This is an
efficiency improvement.

Logrus buffers its output internally, which makes these tests fail
intermittently. They're also not a good example to follow generally.

We now have acceptance tests that exercise this functionality so I'm
pretty relaxed about losing the expectations. However, we can test
them by inspecting the server-received metadata too, so there's no loss
of coverage here.

The move from logrus to labkit for logging also makes these tests hard
to justify keeping.

Gitaly project now properly respects module release flow
and includes a module suffix in the package name. It requires
to re-write all non-suffixed imports with suffixed of a specific
version of tha module. With proper module versioning we don't
need to use a 'replace' directive to point to specific commit
and can use semantic versioning for the gitaly dependency.

Part of: https://gitlab.com/gitlab-org/gitaly/-/issues/3177

Previously, opentracing (if configured) was initialized late in the
gitlab-shell process's lifespan, coming just before making a gRPC
call to Gitaly.

By moving the opentracing initialization to be at process startup, we
make it available for the whole process lifecycle, which is very useful
to gitlab-sshd, as it means we'll only call tracing.Initialize() once
on process startup, rather than once per SSH connection.

To get this working, we need to introduce a context to gitlab-sshd.
This carries the client/service name, but also carries an initial
correlation ID. The main outcome of this is that all calls to the
authorized_keys endpoint from a given gitlab-sshd process will now
share a correlation ID. I don't have a strong opinion about this either
way.

Changelog: fixed