Updates README with improved technical accuracy and examples

[LoserCheems]

Improves technical terminology by replacing "computational efficiency" with "sparse computation capabilities" for better precision.

Updates complexity notation from O(N·k) to O(N·w) to align with implementation variable naming.

Raises minimum requirements to Python 3.8+, PyTorch 2.0+, CUDA 11.8+, and compute capability 8.0+ to reflect actual supported configurations.

Adds comprehensive Quick Start section with complete working code example including proper tensor shapes and device setup.

Expands documentation with detailed benchmarking section covering forward pass equivalence, performance testing, gradient computation, and multi-query associative recall.

Enhances troubleshooting guide with specific commands for verifying CUDA setup, handling import errors, monitoring memory usage, and addressing numerical stability issues.

[Copilot]

Pull Request Overview

This PR refines the README by improving technical terminology, raising dependency requirements, and adding a comprehensive Quick Start guide along with enhanced benchmarking and troubleshooting instructions.

• Replaced generic “computational efficiency” and complexity notation to more precise terms.
• Updated prerequisites to Python 3.8+, PyTorch 2.0+, CUDA 11.8+, and compute capability 8.0+.
• Added a detailed Quick Start example, benchmarking commands, and expanded troubleshooting steps.

Comments suppressed due to low confidence (2)

README.md:72

• The Quick Start imports and uses flash_dma_cuda, but the installation verification later references flash_dma_cpp; unify the module name to avoid confusion.

output = flash_dma_cuda.fwd(

README.md:76

• The variable keep_window_size is not defined before use in the Quick Start snippet; consider declaring it (e.g., keep_window_size = 2048) or replacing it with a literal value.

    keep_window_size=keep_window_size,