Synchronize LLAMA_API with ggml-org/llama.cpp and update cuda workflow for windows

[JamePeng]

Update llama.cpp version llama.cpp updated [from 794fe2 to f08f4b3]
Use the llama_sampler_init instead of llama_sampler() for safe usage
Sync llama : add Phi-4-mini support
Sync llama : expose llama_model_n_head_kv in the API
Sync tool-call: fix Qwen 2.5 Coder support, add micro benchmarks, support trigger patterns for lazy grammars
class LlamaSampler: append add_xtc(), add_top_n_sigma() and add_dry()
Remove Tail-Free sampling
Add TopN-Sigma/XTC/DRY samplers code into sampler
Sync llama : Add Gemma 3 support

[JamePeng]

I tried to adjust the workflow output based on VS2022 to compile pip wheels, and generate two cuda versions 12.4.1 and 12.6.3 and the win version of py310-312 for your convenience.
It should have been compiled now: https://github.com/JamePeng/llama-cpp-python/releases

[JamePeng]

llama.cpp : refactor llama_context, llama_kv_cache, llm_build_context (ggml-org/llama.cpp#12181)
They change API name again, :<

[JamePeng]

The adjusted code is moved to https://github.com/JamePeng/llama-cpp-python/tree/1966-branch