When trying to allocate the username that is required to ask for SSH
authentification methods, we check the return value of `git__strdup` but
will not set up a proper error code. Because of this,
`_git_ssh_setup_conn` may return successfully while there was in fact an
out of memory error.

Fix the issue by calling `git_error_set_oom` and setting the error code
to `-1`.

When creating the reflog message "clone: from %s", we do not check the
call to `git_buf_printf` for an error. As we use `git_buf_cstr`
immediately afterwards, we may thus try to use the `git_buf_oom` buffer
without noticing. While this is unproblematic due to the buffer
containing a single NUL byte, only, we should certainly notice
out-of-memory situations early.

Fix this by checking the return value of `git_buf_printf`.

In order to be able to verify that SSH authentication with user and
password works correctly, we need to first set up a password and make it
available via the GITTEST_REMOTE_PASS environment variable.

The test suite online::clone::env will by default refuse any
certificates presented by the remote configured via environment
variables. In order for the tests to succeed with our instance of sshd,
we should either make the expected host key available via the
environment or alternatively just override it by setting
GITTEST_REMOTE_SSL_NOVERIFY.

This commit goes the easier way of just overriding the check, as the
point of the whole series is to first get the online::clone test suite
into a state where it's much easier to improve it at a later point in
time.

The libssh2 library may leak memory when performing the session
handshake in various edge cases, like e.g. timeouts. As there is
multiple memleaks that may happen, depending on when the exact failure
happens inside of libssh2_session_handshake, we shouldn't specify a
detailed stack of which memleaks to suppress, but just suppress all that
happen with `libssh2_session_handshake`.

The clone::nonetwork test suite still has a define for a GitHub
repository URL, which naturally is not tested in the no-network test
suite anymore. Remove it.

Modernize coding style of the clone test suite. Most importantly, we
introduce defines for each of the repo URLS and consistently rely on the
cleanup function to remove the sandbox repository for us.

The test suite online::clone is quite huge and hard to grasp. It
intermingles tests for HTTP and SSH-based clones as well as the ones
that make use of a server defined via environment variables. Split these
all up into different tests to make the suites more focused as well as
easier to understand. Furthermore, this will allow us to introduce
variants for our HTTP tests to more thoroughly test against different
external hosting services.

Right now, we only run each clone test against a single Git provider,
making it hard to understand which features are tested against which
provide. Furthermore, more intricate behaviour like e.g. authentication
is thus not thoroughly tested against the different implementations,
which may differ slightly across hosters.

Introduce test variants to start fixing this. For each provider, we will
initialize an optional set of URLs to test against where each URL aims
to exercise some specific behaviour. For now, we implement tests against
GitHub, Bitbucket, Azure and GitLab. Tests against GitLab and Azure are
completely new.

The variants also uncovered behavioural inconsistencies in how Bitbucket
handles nonexistent repositories: instead of trying to authenticate, it
will directly return a 404 status code and thus reveal that no such repo
exists, not even in private. We thus special-case this behaviour in our
test suite.