This reverts commit 091b8d58b3502a99b3bfbec9db457c92cc736beb, reversing
changes made to 9b2ca9cf8e.

This reverts commit b4b3ea9d99.

Just merge cuda and avx detection and change layout.

had to update binary to `b1426`

CPU Feature Detection

Based on this version: 6b73ef1201

- Minor changes to cleanup some of the compiler warnings

- Added some more comments
 - Modified `llama_tokenize` to not allocate
 - Modified `llama_tokenize_native` to take a pointer instead of an array, allowing use with no allocations
 - Removed GgmlInitParams (not used)

- Added disposal check into `SafeLLamaContextHandle`

- Also refactored it to return an `int[]` instead of an `IEnumerable<int>`, solving the "multiple enumeration" problems at the source!

This change _only_ implements the low level API and makes no effort to update the LlamaSharp higher level abstraction.

It is built upon llama `b3f138d`, necessary DLLs are **not** included in this commit.

Memory Mapped LoadState/SaveState

- A C# array cannot exceed 2,147,483,591 bytes. In my own use of LlamaSharp I encountered this limit.
 - This saves an extra copy of the entire state data into a C# `byte[]`, so it should be faster.

This does _not_ fix some other places where `GetStateData` is used. I'll look at those in a separate PR.

Adapt the interface change in llama_backend_init