Min et al. variant of in-context learning

[dirkgr]

Motivation: It's a good baseline that should be easy to implement in the catwalk context, but nobody has asked for it.

Described by Liu at al like this: Min et al. [21] proposed ensemble ICL, where instead of using the output probability from concatenating the k training examples, the output probabilities of the model on each training example (i.e. 1-shot ICL for each of the k examples) are multiplied together. This lowers the memory cost by a factor of k/2 but increases the computational cost by a factor of 2. In terms of task performance, Min et al. [21] find that ensemble ICL outperforms the standard concatenative variant.

This depends on first getting normal few-shot ICL working on Catwalk.