Add on-paper form of RoPE kernel

[Comet0322]

Summary

Implement the on-paper form of the RoPE kernel from RoFormer. 
This implementation does not support optional value input, unlike the HuggingFace RoFormer RoPE implementation.

Details

The code is adapted from Liger Kernel's RoPE implementation.
In the current Liger Kernel's RoPE implementation, the head is divided into left and right parts for computation:

$y = [x1, x2] * [cos, cos] + [-x2, x1] * [sin, sin]$
$dy = [dx1, dx2] * [cos, cos] + [-dx2, dx1] * [-sin, -sin]$

Corresponds to the vector-vector multiplication-addition form:

$$\begin{pmatrix} q_0\\\ q_1\\\ q_2\\\ \vdots\\\ q_{d/2-1}\\\ q_{d/2}\\\ q_{d/2+1}\\\ q_{d/2+2}\\\ \vdots\\\ q_{d-1} \end{pmatrix} \otimes \begin{pmatrix} cos\ m\theta_0 \\\ cos\ m\theta_1 \\\ cos\ m\theta_2 \\\ \vdots\\\ cos\ m\theta_{d/2-1} \\\ cos\ m\theta_0 \\\ cos\ m\theta_1 \\\ cos\ m\theta_2 \\\ \vdots\\\ cos\ m\theta_{d/2-1} \end{pmatrix} + \begin{pmatrix} -q_{d/2}\\\ -q_{d/2+1}\\\ -q_{d/2+2}\\\ \vdots\\\ -q_{d-1}\\\ q_0\\\ q_1\\\ q_2\\\ \vdots\\\ q_{d/2-1} \end{pmatrix} \otimes \begin{pmatrix} sin\ m\theta_0 \\\ sin\ m\theta_1 \\\ sin\ m\theta_2 \\\ \vdots\\\ sin\ m\theta_{d/2-1} \\\ sin\ m\theta_0 \\\ sin\ m\theta_1 \\\ sin\ m\theta_2 \\\ \vdots\\\ sin\ m\theta_{d/2-1} \end{pmatrix}$$

To obtain the on-paper form of RoPE, the head is instead divided into even-indexed and odd-indexed parts for computation:
$y_{even} = x_{even} * cos - x_{odd} * sin$
$y_{odd} = x_{odd} * cos + x_{even} * sin$

$dy_{even} = dx_{even} * cos + dx_{odd} * sin$
$dy_{odd} = dx_{odd} * cos - dx_{even} * sin$

Corresponds to the vector-vector multiplication-addition form:

$$\begin{pmatrix} q_0\\\ q_1\\\ q_2\\\ q_3\\\ \vdots\\\ q_{d-2}\\\ q_{d-1} \end{pmatrix} \otimes \begin{pmatrix} cos\ m\theta_0 \\\ cos\ m\theta_0 \\\ cos\ m\theta_1 \\\ cos\ m\theta_1 \\\ \vdots\\\ cos\ m\theta_{d/2-1} \\\ cos\ m\theta_{d/2-1} \end{pmatrix} + \begin{pmatrix} -q_1\\\ q_0\\\ -q_3\\\ q_2\\\ \vdots\\\ -q_{d-1}\\\ q_{d-2} \end{pmatrix} \otimes \begin{pmatrix} sin\ m\theta_0 \\\ sin\ m\theta_0 \\\ sin\ m\theta_1 \\\ sin\ m\theta_1 \\\ \vdots\\\ sin\ m\theta_{d/2-1} \\\ sin\ m\theta_{d/2-1} \end{pmatrix}$$

Testing Done

• Hardware Type: A100-80G-PCIe
• run make test to ensure correctness
• run make checkstyle to ensure code style
• run make test-convergence to ensure convergence

[Comet0322]

I've added a paper-form option to the current Liger Kernel RoPE implementation.