Fix FlexAttention cross-attention masks for dynamic text lengths

[Arison591]

Summary

• derive the per-sample text sequence length from the projected text embeddings
• pass that length into FlexAttnFunc.init_mask
• build the cross-attention sequence IDs from the runtime length instead of a fixed 512 tokens

Root cause

During post-training, text embeddings are flattened from [B, L, C] to [1, B * L, C] before cross-attention. The FlexAttention cross mask was always built for B * 512 key/value tokens, so any embedding length other than 512 produced a mask/KV shape mismatch.

This keeps the existing per-sample masking semantics while making the mask length match the actual projected text sequence.

Fixes #82.

Validation

• NVIDIA GeForce RTX 5090 (compute capability 12.0)
• PyTorch 2.9.0+cu128 and Triton 3.5.0
• reproduced the old behavior with B=2, L=20: mask shape (1, 1, 256, 1024) rejected runtime kv_len=40
• validated the patched FlexAttnFunc with B=2, L=20: mask shape (1, 1, 256, 40), BF16 output shape (1, 256, 2, 64), all values finite
• python -m py_compile wan_va/modules/model.py
• git diff --check