The current behavior of the llama.cpp LLM pipeline is to always set n_ctx=0. When n_ctx=0, the context size defaults to n_ctx_train which can be very large with some models.

This change will fallback to the default n_ctx when n_ctx=0 fails due to being out of memory. It will also allow n_ctx as a input parameter. If a manually set n_ctx is too large, this will fail since it's user-specified.