Diffusion Fine-Tuning via Reparameterized Policy Gradient of the Soft Q-Function