Update integration

[LoserCheems]

No description provided.

[Copilot]

Pull Request Overview

This PR restructures and streamlines the integration guide for combining Flash Attention with Dynamic Mask Attention. It provides a concise overview, updates the table of contents, and reorganizes key sections including architecture, core modifications, performance, and API changes.

• Revamped title, overview, and high-level description
• Added new, focused Table of Contents and reorganized section layout
• Updated and detailed API changes, performance considerations, and memory/layout sections

Comments suppressed due to low confidence (2)

docs/integration.md:13

• The ToC entry "Implementation Details" doesn’t match any H2 heading; consider adding or renaming the corresponding section header to ensure the link works.

3. [Implementation Details](#implementation-details)

docs/integration.md:7

• [nitpick] The term "Zero-Order Hold states" (ZOH) is used here without defining the acronym; consider expanding and defining ZOH on first use for better reader clarity.

The integration implements a two-stage approach: Python frontend pre-computes Zero-Order Hold states and Active Mask tensors, while the CUDA backend performs sparse attention computation using these pre-computed masks.