This adds a new `auth` keyword param to `Net::SMTP.start` and `#start`
that can be used to pass any arbitrary keyword parameters to
`#authenticate`.  The pre-existing `username`, `secret`, etc keyword
params will retain their existing behavior as positional arguments to
`#authenticate`.

Although "user" is a reasonable abbreviation, the parameter is more
accurately described as a "username" or an "authentication identity".
They are synonomous here, with "username" winning when both are present.

Username can be set by args[0], authcid, username, or user.
Secret can be set by args[1], secret, or password.

Since all of the existing authenticators have the same API, it is
sufficient to update `check_args` in the base class.

This API is a little bit confusing, IMO.  But it does preserve backward
compatibility, while allowing authenticators that don't allow positional
parameters to work without crashing.  But, authenticators that require
only one parameter—or more than three—will still be inaccessible.

This is convenient for `smtp.start auth: {type:, **etc}`.

Although `#authenticate` can be updated to make username and secret
_both_ optional, by placing the mechanism last and making it optional,
it's not possible to use an authenticator with a _single_ positional
parameter or with more than two positional parameters.  By placing
`type` first among positional parameters or as a keyword argument, we
avoid this problem.

The net-imap dependency requires 2.7.3, to deal with kwargs.

This commit adds the `net-imap` as a default fallback for mechanisms
that haven't otherwise been added.  In this commit, the original
implementation is still used by `#authenticate` for the `PLAIN`,
`XOAUTH2`, `LOGIN`, and `CRAM-MD5` mechanisms.  Every other mechanism
supported by `net-imap` v0.4.0 is added here:
* `ANONYMOUS`
* `DIGEST-MD5` _(deprecated)_
* `EXTERNAL`
* `OAUTHBEARER`
* `SCRAM-SHA-1` and `SCRAM-SHA-256`

**TODO:** Ideally, `net-smtp` and `net-imap` should both depend on a
shared `sasl` or `net-sasl` gem, rather than keep the SASL
implementation inside one or the other.  See
ruby/net-imap#23.

This is mostly to stay consistent with other SASL mechanisms and to
remove duplication:  We shouldn't have unique SASL mechanism
implementations for each individual protocol client.

In some cases, the versions from `net-imap` are more full-featured.
E.g: they support `authzid` and print deprecation warnings for MD5
authenticators.

The tests are updated to silence deprecation warnings, which are printed
by the authenticators from `net-imap`.

The XOAUTH2 authenticator that comes from Net::IMAP::SASL does support
keyword attributes.  But XOAUTH2 is very simple, so this doesn't really
add much.

This is mostly to stay consistent with other SASL mechanisms and to
remove duplication:  We shouldn't have unique SASL mechanism
implementations for each individual protocol client.

This renames `#do_finish` to `#quit!`, makes it public, adds an option
to convert exceptions into warnings, and extracts the contents of the
ensure block into a new public `#disconnect` method.  The motivation for
making these public is given in the rdoc.

As documented in the rdoc, `#quit!`:
> Calls #quit and ensures that #disconnect is called.  Returns the
> result from #quit.  Returns +nil+ when the client is already
> disconnected or when a prior error prevents the client from calling
> #quit.  Unlike #finish, this an exception will not be raised when the
> client has not started.
>
> When <tt>exception: :warn</tt> is specified, when #quit raises a
> StandardError, a warning will be printed and the exception is
> returned, not re-raised.  When <tt>exception: false</tt> is specified,
> a warning will not be printed.  This is useful when the connection
> must be dropped, for example in a test suite or due to security
> concerns.

As documented in the rdoc, `#disconnect`:
> Disconnects the socket without checking if the connection has started
> yet, and without sending a final QUIT message to the server.
>
> Generally, either #finish or #quit! should be used instead.