[Perf] flashdreams.accelerated Library #482

Issue Description

FlashDreams currently expose several levels of acceleration mechanism:

• CUDA graph capture
• Inference prewarm
• Native acceleration kernel

While CUDA graph and prewarm are built into flashdreams/flashdreams/infra/acceleration, the native acceleration kernels are optimized per integration.

Proposed Solution

Build a flashdreams.accelerated (later refactored into flashdreams.core) library that implements generic PyTorch module and Triton kernels for accelerated inference.

The library should reside in flashdreams/flashdreams/accelerated (later refactored into flashdreams/core/module)

The library should provide:

• flashdreams/accelerated/
  • triton/: Triton kernels (flashattention2, sageattention3, fused norm + rope + kv update, etc.).
  • cuda/(optional): highly specialized CUDA kernel.
  • PyTorch modules that uses these accelerated kernels (multi head attention, quantized linear, etc.)