llama : add check for KV cache shifts

[ggerganov]

fix #10380

Disallow context shifts for models that do not support it (such as DeepSeek V2). Add

bool llama_kv_cache_can_shift(struct llama_context * ctx);

[slaren]

Should this return false for recurrent models as well? Not sure what's the logic there, but llama_kv_cache_update_internal silently ignores models with LLAMA_ROPE_TYPE_NONE.

[ggerganov]

Yes, it's likely needed to return false for recurrent models.

The reason to do nothing in llama_kv_cache_update_internal when the rope type is none is because when we apply shifts to the KV cache using functions like llama_kv_cache_seq_add(), we do 2 things:

• Update the positions of the KV cells - i.e. just modify the meta data in llama_kv_cell
• Re-rope the data in the KV cells

The later step is necessary only if the data is roped. For ALiBi models for example, we should not apply this second step, but in theory we still support "shifting" the KV cache for those models, since the positional information is in the KQ mask.