Add regression test for MusicgenMelody audio conditioning (GH #45647)

[voodoovampire]

This PR adds a regression test for issue #45647.

Test behavior:

• Creates two different 1-second sine reference audios (A4 at 440 Hz, Eb4 at 311.13 Hz).
• Generates audio from each using AutoProcessor with text=["jazz"], same seed, same params.
• Asserts that the generated outputs differ by at least 1e-3 in mean absolute difference.

Current status:
On facebook/musicgen-melody, this test fails with mean abs diff=0.00000000, confirming that audio conditioning is ignored.

Next step:
This test provides a reliable repro. The model fix is being coordinated with @adityachoksi2512, who has already bisected the regression to the cache refactor in PR #38635.

Closes #45647 (once the model fix is merged).

[adityachoksi2512]

Hey @voodoovampire - thanks for the co-author credit! Could you update the email in the co-authored-by line to adityachoksi2512@gmail.com? The current one is an old work email that won't link to my GitHub profile. Also looks like there's a duplicate entry with a slightly different username. Thanks!

[voodoovampire]

Hey bro @adityachoksi2512 - absolutely! Just updated the co-authored-by email to adityachoksi2512@gmail.com. The duplicate was my mistake - I initially used the work email from your original commit metadata. Should be fixed now!
Thanks for being collaborative about this. Really appreciate it. 🙏

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: musicgen_melody