Implement Variants of RoPE

[ruanjm]

ATT. Includes:

• Rotate style: Supports both NEOX and GPT-J style. Backward is also supported by both styles.
• The size in hidden dim of freqs/sin/cos can be half of that of input/output tensor.
• Nope first: Rotate latter half of tensor.
• 2 channels: handle two inputs at once.
• Inplace mode: Input and output is the same. Some specific optimization is also designed for this mode.
• Get rotate angle or cos/sin via positions and offsets.

Compared with legacy implementation, the average latency in fp16 and bf16 is reduced to about 73.5% (from 37.1% in best cases to 142.5% in worst cases). Meanwhile, more functions are supported.