Efficient Training of Self-Supervised Speech Foundation Models on a Compute Budget