Fix LTXVAudioVideoMask negative mask padding

[Booyaka101]

Fixes #615.

LTXVAudioVideoMask crashes with Trying to create tensor with negative dimension when an existing noise_mask has more frames than the padded samples.

In the existing-mask pad branch the guard checks that the samples grew (samples.shape[2] > original_samples.shape[2]), but the padding amount is samples.shape[2] - mask.shape[2]. When the incoming noise_mask already has more frames than the padded samples, that subtraction is negative and torch.zeros(..., negative, ...) raises. The reporter's [1, 1, -32, 640] is the audio branch (4D); the video branch (5D) has the same defect.

Fix: compute extra = samples - mask and pad only when it is positive, otherwise trim the mask to the sample length (matching the existing truncate behaviour). Applied to both branches.

Verified with a CPU repro of the exact shape arithmetic (samples padded to fewer frames than the existing mask):

before: RuntimeError: zeros: Dimension size must be non-negative
after:  samples=(1, 128, 6, 640) mask=(1, 1, 6, 640)  (mask frames == samples frames)