optimize inference performance

[zhuxiaoxuhit]

• odeint_euler: only keep final state instead of stacking full trajectory, saves O(steps * batch * seq * dim) memory
• clone x in ode function to avoid mutating integrator state in-place
• fix _project() using .to(device_type) which loses cuda device index on multi-gpu
• move einops import to module level
• use nn.Sequential for attention to_out
• reuse encode_prompt_audio() result to avoid duplicate vae encoding and duplicate audio loading in inference scripts
• add prompt_latent / prompt_duration_frames params to forward() for pre-encoded prompt