Feature: Add Fused Attention Kernel

[ShlokVFX]

🚀 Feature: Fused Attention Kernel

Overview

Implement a fused attention kernel to improve performance by combining
QKV projection, attention score computation, softmax, and output projection
into a single optimized kernel.

Motivation

• Reduce memory bandwidth overhead
• Minimize kernel launch latency
• Improve end-to-end attention throughput
• Align with FlashAttention-style optimizations

Scope

• Single-GPU support initially
• FP16 / BF16 (FP8 optional later)
• CUDA kernel (Triton or CUDA C++)
• Drop-in replacement for existing attention path

Planned Approach

• Implement fused QKV + attention + output
• Use shared memory / SRAM tiling
• Optimize for RTX 3060 (SM86)
• Benchmark against baseline attention

Acceptance Criteria

• Correctness matches reference attention
• Lower latency than unfused implementation
• Reduced memory reads/writes
• Benchmark CSV generated

Out of Scope (for now)

• Multi-GPU
• Backward pass
• Training support

Notes

This is a planned feature and will be implemented in a future PR.