Build the linear noise schedule of a diffusion model with num_steps timesteps and return its three per-step arrays.

Here $T$ is num_steps, the timestep $t$ runs from 0 to $T - 1$, $\beta_{\text{start}}$ is beta_start and $\beta_{\text{end}}$ is beta_end. The betas are evenly spaced, so the first one equals beta_start and the last one equals beta_end. Each $\alpha_t$ is the fraction of variance kept at step $t$, and each $\bar\alpha_t$ is the running product of all alphas up to and including step $t$. When num_steps is 1 the schedule holds the single value beta_start.

Return a dict with exactly the keys "betas", "alphas" and "alpha_bars", where each value is a float64 NumPy array of shape (num_steps,) holding $\beta_t$, $\alpha_t$ and $\bar\alpha_t$ in timestep order.