New 'Scatter' sampler added to llama.cpp

A new experimental sampler called Scatter has been implemented in llama.cpp. This sampler finely adjusts the probability mass among the top candidates within the model's next-token probability distribution, preventing the generated output from becoming overly rigid.

The existing Temperature adjustment method smooths the entire distribution, allowing low-probability 'tail' tokens to participate in generation, which can lead to inappropriate word choices or context deviation. In contrast, Scatter has the following characteristics.

• Local Smoothing: Exchanges probability only between adjacent ranks within the selected top K candidates.
• Tail Distribution Protection: Smooths only the head of the probability distribution, preventing probability from leaking into the low-probability tail.
• Key Features: Includes fixed strength adjustment, adaptive strength adjustment via entropy feedback, and repeated token absorption functionality.

Currently, this feature is implemented as an experimental sampler, and while it is included within the default sampler chain, it is disabled (noop) by default.