Reconcile DDIM scheduler behavior, tests, and comments

[blacks1ne]

Summary

The DDIM implementation, inline comments, and unit-test expectations disagree, causing two default diffusion tests to fail.

Reproduction

cargo test -p klearu-diffusion

Current failures:

• scheduler::ddim::tests::timesteps_descending
• scheduler::ddim::tests::step_at_t0_returns_x0_pred

Evidence

The implementation now targets Diffusers-style timestep_spacing="leading", steps_offset=1, and set_alpha_to_one=false. The tests still expect the previous [980, ..., 0] schedule and final alpha of 1.0. A nearby comment says alpha_bar_at(-1) = 1.0, while the function returns alpha_bars[0].

Proposed resolution

Compare one fixed scheduler configuration and reference vector against Python Diffusers, then update the implementation, tests, and comments together.

Acceptance criteria

• cargo test -p klearu-diffusion is green.
• Tests lock the intended timestep sequence and final-step behavior to a reference.
• Comments accurately describe steps_offset and set_alpha_to_one semantics.