Fix NotImplemented errors, xformers attention shape, and missing text conditioning

[Mr-Neutr0n]

Summary

• raise NotImplemented -> raise NotImplementedError in models/latte.py (line 73) and models/latte_img.py (line 76). NotImplemented is a special singleton used for binary operator fallbacks, not an exception class. Using it with raise produces a TypeError instead of the intended error, masking the real issue.

• Transpose q, k, v for xformers attention in models/latte_img.py (line 61). After permute(2, 0, 3, 1, 4) and unbind, tensors are shaped (B, heads, N, dim), but xformers.ops.memory_efficient_attention expects (B, N, heads, dim). The implementation in models/latte.py (lines 55-58) correctly transposes before calling xformers; this patch applies the same fix to latte_img.py.

• Add missing elif self.extras == 78 before final layer in models/latte.py (line 372). The temporal block loop correctly handles extras == 78 by adding text_embedding_temp to the conditioning, but the final adaptive layer norm block only checked for extras == 2 (class conditioning) and fell through to unconditional for all other values. This meant text-conditioned generation (extras == 78) silently dropped text conditioning at the final layer.

Test plan

• Verify raise NotImplementedError is correctly raised when an unsupported attention mode is passed
• Run inference with xformers attention mode using latte_img.py and confirm no shape errors
• Run text-conditioned video generation and verify text conditioning is applied through the final layer