Fix 24486: TP: allows the usage of 4-10 gpus for stepfun

[krampenschiesser]

Overview

This fixes #24486 and allows the usage of 4-10 and probably more gpu's. I cannot test with <4 because of too little vram.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes I used llama.cpp and qwen3.5 122b as an advanced grep tool for discovering the codebase all code comes from me (as little as there is).

[krampenschiesser]

I am seeing some problems with tensor splits by going larger with granularity. I will try to implement the discovery of additional tensors during the std::vec initialization. this will include the passthrough of the user data+model into the ggml_backend_meta_get_split_state function so I can ask the model about additional tensors by tensor name.
will post some logs tomorrow regarding the core issue with the graph.

[krampenschiesser]

This is way better.
works in all configurations i could test (4-10 gpus) probably with 2,3 and 10+ too.
Tensor splits look better too.
Should be good for review.
@JohannesGaessler please take a look, i don't know if it is ok to add more model specific exceptions into the llama-model.cpp