llama : merge conts and reshapes and remove unnecessary cont

[CISC]

Merged all cont+reshape_*d into cont_*d and remove unnecessary ggml_cont in mtmd.

[CISC]

Note: I did see this #13097 (comment) about the ggml_cont I removed, but the issue has either since been resolved, or it was just a mix-up during implementation refactoring. I have tested Pixtral (on CUDA) with this change, and it works fine.

[CISC]

@ggerganov Hmmm, two 124 failures in a row for ggml-100-mac-m4:
https://github.com/ggml-org/ci/tree/results/llama.cpp/ba/a9255a45105d2d3b4ec432af13b7a6eda3ff35/ggml-100-mac-m4

I can't find where it fails though, but seems to be related to Pythia tests...

[CISC]

@ggerganov Hmmm, two 124 failures in a row for ggml-100-mac-m4: https://github.com/ggml-org/ci/tree/results/llama.cpp/ba/a9255a45105d2d3b4ec432af13b7a6eda3ff35/ggml-100-mac-m4

I can't find where it fails though, but seems to be related to Pythia tests...

Looking at the stdall log it looks like it dies right at the point where it downloads wikitext, so probably not an actual issue with the test. Weird that it's only failing for ggml-100-mac-m4 though?

[ggerganov]

I'm looking into this - thanks for noticing. It's not related to the code, just the internet connection issues.