ggml : sync llama.cpp (#409)

ggml : sync llama.cpp (#409)

* ggml : sync llama.cpp

ggml-ci

* ggml : fix nullptr derefs in backward

* ci : add mnist test, import/export graph

* add op_params to ggml_graph_export/import

ggml-ci

* mnist : export/import op_params for testing purposes

* mnist : fix f32 model generation test + instructions

ggml-ci

* ci : install python deps even for low-perf builds

ggml-ci

---------

Co-authored-by: Diego Devesa <redacted>