Find common ground between what Jan is doing and what's happening in ggml with Training

[janchk]

Let's start looking at what's happening over here https://github.com/ggerganov/ggml/tree/master/examples/mnist down to the flow of PRs and the kinds of discussions folks are having there (e.g. ggml-org/ggml#982).

The work being done there is obviously a massive refactoring of an earlier attempt at llama.cpp finetuning that was ripped out some time ago waiting for the above ggml research to conclude. The previous work was around here ggml-org/llama.cpp#8669 and here ggml-org/llama.cpp#2632)

It may actually make sense to look back at the last release of llama.cpp that still had that functionality just to appreciate the scope a little bit.

[janchk]

Scope of work that needed to be done:

1. Implement noisy multiplication operation like GGML_OP_MUL_MAT
   https://github.com/ggerganov/ggml/blob/6dccc647264f5429df2624f36138f601e7ce23e5/src/ggml-backend.cpp#L1186 // Only for CPU without BLAS, etc.
   https://github.com/ggerganov/ggml/blob/6dccc647264f5429df2624f36138f601e7ce23e5/src/ggml.c#L5635
2. Implement potenial loss
   https://github.com/ggerganov/ggml/blob/6dccc647264f5429df2624f36138f601e7ce23e5/examples/mnist/mnist-common.cpp#L706
3. Implement adjustable LR (hardcode?)
4. Implement mixed noise scaler (hadcode?)
5. Implement graph analyzer to gather weight noise scales (hardcode?)
6. Implement distillation techinque

[janchk]

https://github.com/aifoundry-org/ggml/tree/dev/noisy_mul
Implemented noisy multiplication.
Further development require either deep talk with llamacpp contributor (or help from him) or large amount of time.