[TASK] Add a new NLI ball for non-linear operators

[shirohasuki]

Deliverables

• Add a Nonlinear-Operators‘ RTL implementation in the prototype lib (under the arch path).
• A Pull Request (PR) containing a test written in C for the Common Non-linear operation and a README to introduce your design.
• Report the performance results in this issue.

Task Description

• LLMs rely heavily on nonlinear operators like SiLU, RMSNorm, and Softmax. While linear layers are highly optimised, these nonlinear functions often require expensive high-precision floating-point arithmetic. Traditional lookup table (LUT) methods often fail to handle the extreme outliers common in modern LLMs, leading to significant accuracy degradation.
• NLI is a calibration-free, hardware-friendly framework designed to efficiently approximate different nonlinear functions. Offline find optimal cutpoints, NLI maintains high accuracy even for extreme outliers in LLM inference, where traditional LUT methods usually collapse.
• You can learn this methodology from the paper, "NLI:Non-uniform Linear Interpolation Approximation of Nonlinear Operations for Efficient LLMs Inference". Design a ball for Buckyball based on this methodology.
• Please refer to the previous Pull Request (Completed the development of ReluBall and further improved the operation manual #6) for the detailed ball implementation.