memory: add is_iswa for memory_hybrid

[ngxson]

Alternative to #18601 , reuse llama_memory_hybrid instead of duplicating the class

(I haven't tested it yet, just pushing this PR for discussion)

[ggerganov]

I prefer the #18601 approach. Indeed, it is duplicating significant chunk of the code, but here we still have to add the new llm_graph_input_mem_hybrid_iswa, or alternatively - add the extra is_iswa logic there in a similar way. On the other hand, the logic of each class in #18061 is more straightforward and easier to follow.

My recommendation is to proceed with llama_memory_hybrid_iswa for now.

[ngxson]

Got it, thanks! let's go with the other PR then

[ggerganov]

There might be a simplification where instead of having both llama_kv_cache and llama_kv_cache_iswa, we keep just llama_kv_cache_iswa since it is technically a superset of llama_kv_cache functionality. But we can do this simplification after this.