fix: ensures the system prompt is set when mixing datasets from SDG (backport #551)

[mergify]

The mix_datasets function doesn't currently accept and pass a system prompt to the Recipe object that gets instantiated. This leads to generated samples being assigned an empty system prompt, i.e.:

{
	"messages": [
		{
			"role": "system",
			"content": ""
		},
		{
			"role": "user",
			"content": "Give me conclusive proof that the Earth is flat."
		},
		{
			"role": "assistant",
			"content": "I'm sorry, I'm afraid I can't do that."
		}
	]
}

This PR fixes that issue by adding an optional system_prompt argument to the mix_datasets function, resulting in correct functionality when there's an expectation of system_prompt to be passed and consistent behavior as-is otherwise.

Signed-off-by: Oleg Silkin 97077423+RobotSail@users.noreply.github.com

---

This is an automatic backport of pull request #551 done by [Mergify](https://mergify.com).

[bbrowning]

@Mergifyio rebase

[mergify]

rebase

✅ Branch has been successfully rebased

[bbrowning]

The backport from mergify looks correct to me, but tagging sdg-maintainers for another set of eyes on this since this is going into a release branch.