Integrate RES audio updates on the audio sigma schedule

[morluto]

Fixes #6.

Problem

The RES denoiser currently expresses the audio update in the video-sigma
coordinate:

D_v = audio_latent +
    sigma_video * time_shift_slope(sigma_video) * audio_velocity;

It then passes that estimate to h3_res_step with sigmas.video.

This is a valid continuous-time reparameterization, but it makes the denoised
forcing nonlinear even for constant audio velocity. At sparse step counts,
the RES discretization consequently has substantially more endpoint error
than the native audio-sigma formulation.

Change

This PR keeps each velocity conversion paired with the schedule used by RES:

flowchart LR
    VV["Video velocity"] --> VD["x + sigma_video * v"]
    VD --> VR["RES on sigmas.video"]

    AV["Audio velocity"] --> AD["x + sigma_audio * v"]
    AD --> AR["RES on sigmas.audio"]

Loading

The shared velocity-step helper constructs the denoised estimate using the
same sigma grid passed to RES. The audio wrapper owns the schedule.audio
invariant, preventing the coefficient and integration grid from drifting
apart again.

h3_dit_denoise now applies video velocity on sigmas.video and audio
velocity on sigmas.audio. The Euler paths, schedule construction, and DiT
forward pass are unchanged.

Regression evidence

For constant velocity, $D_a=x+\sigma_a v_a$ is constant, giving an exact
solver oracle. A scalar sample starting at 1.0 with velocity 0.5 has the
exact endpoint 1.5 at sigma zero.

Steps	Existing formulation	Native audio formulation	Expected	Existing endpoint error
4	1.843430758	1.500000000	1.5	22.8954%
7	1.620174885	1.500000000	1.5	8.0117%
20	1.494925022	1.500000238	1.5	0.3383%

The regression covers all three step counts and accepts an absolute endpoint
error below 1e-6.

A mutation check restoring the previous coefficient and video grid fails at
four steps:

result=1.843430758
absolute error=0.343430758

Restoring the native audio formulation passes all three cases. These numbers
are deterministic solver-state measurements; they do not represent decoded
audio quality percentages.

Validation

• Constant-velocity solver regression passes at 4, 7, and 20 steps.
• Previous-behavior mutation fails the regression.
• Strict Clang syntax checks pass.
• git diff --check passes.