move STA_configuration.py to fastvideo/attention/backends (#856)

Author: H1yori233

csrc/attn/sliding_tile_attn/README.md

@@ -72,6 +72,15 @@ We give a demo for 2D STA with window size (6,6) operating on a (10, 10) image.
 
 https://github.com/user-attachments/assets/f3b6dd79-7b43-4b60-a0fa-3d6495ec5747
 
+
+## STA Configuration Logic
+Here is a diagram of how the window is configured and passed through the FastVideo pipeline:
+
+<div align="center">
+<img src="../../../docs/source/_static/images/STA_configuration.png" width="80%"/>
+</div>
+
+
 ## Why is STA Fast?
 2D/3D Sliding Window Attention (SWA) creates many mixed blocks in the attention map. Even though mixed blocks have less output value,a mixed block is significantly slower than a dense block due to the GPU-unfriendly masking operation. 
 

docs/source/_static/images/STA_configuration.png

@@ -575,7 +575,7 @@ def prepare_sta_param(self, batch: ForwardBatch,
             fastvideo_args: The inference arguments.
         """
         # TODO(kevin): STA mask search, currently only support Wan2.1 with 69x768x1280
-        from fastvideo.STA_configuration import configure_sta
+        from fastvideo.attention.backends.STA_configuration import configure_sta
         STA_mode = fastvideo_args.STA_mode
         skip_time_steps = fastvideo_args.skip_time_steps
         if batch.timesteps is None:
@@ -682,7 +682,7 @@ def save_sta_search_results(self, batch: ForwardBatch):
             raise NotImplementedError(
                 "STA mask search is not supported for this resolution")
 
-        from fastvideo.STA_configuration import save_mask_search_results
+        from fastvideo.attention.backends.STA_configuration import save_mask_search_results
         if batch.mask_search_final_result_pos is not None and batch.prompt is not None:
             save_mask_search_results(
                 [