Side-car dev container features

This proposal is an enhancement idea to the current [dev container features proposal](https://github.com/devcontainers/spec/blob/main/proposals/devcontainer-features.md).

---

While Features provide a useful way to encapsulate dev container metadata and an installation script into an easily referenced, atomic unit, using a script is not always the ideal way to make something available into a dev container. There are **nuances in getting more complex software running in a container**, and the current mechanism most used to **deliver these best practices** to customers is through a dedicated **container image**. Examples include databases, things like a Redis cache, and other related services a development environment relies on.

If we think about these kinds of services as potential "Features", they are a **different class of feature than a runtime or SDK** would be - they are standalone servers that you interact with "remotely" instead. Downloading container images can also be substantially faster than full installation steps for a typical Linux install if you were to add it into a container. Worse yet, these full installers can intermix with the OS in a way that prevents them from just being "untar'd" into a folder, which limits caching beyond built-in image layering (which only work if the parent layer has not been altered for the same image). Furthermore, this class of images is typically produced by the same **team that maintains the software** and is specifically optimized for use as a **side-car** - whether in dev using something like Docker Compose locally or in production via Kubernetes at deployment time. Thankfully, dev containers are well-suited to work with these types of side-car container images and many dev container definitions/templates encourage their use via Docker Compose.

However, particularly if you are not directly container deploying the application, what developers often expect is to be able to **start working with these services as if they were spun up on their machine on localhost**. Instructions and tutorials often use localhost as the host name, for example. Even when the application is container deployted, the containerization may happen later - either sometime later in a given project/app's lifetime or just automatically as a part of a later stage in the flow to production. Furthermore, even with side-car containers, you may **still need or want utilities** to be available in your primary container (e.g., a Mongo shell or Postgres command line tools) so that you can interact with the service from the command line. While these tools are a fast install, it **separates out the configuration into two steps**.

To address this issue, we could enhance the concept of a Dev Container Feature to **not only provide scripts** for a referneced container, but also **optionally spin up a separate image pre-configured to operate on the same network**. Orchestrator formats like Compose continue to be the core mechanism for general orchestration tasks, but developers can then add a side-car container with a single line reference. The combination of dev container metadata, a script, and a separate running container on the same network should then give a "**local feel**" but with **substantially less Feature code and a faster overall install time**.

This is a topic that has been raised in a few different spots and issue repos, so I'm raising this to track it as a possible addition to the spec.