Update ggml

[saharNooby]

Notable improvements:

• sequence lengths of up to 64 are now allowed to be used out-of-the-box, without the need to manually change the node limit in ggml.h
• a small performance boost

k-quants support was NOT added since there is no support yet for them in the main ggml repo.

On the internal side:

• "future tensors" mechanism was removed in favor of new way of calculating required context size: a measuring allocator
• a couple of new tests were added, some existing ones were fixed