Major llama.cpp API Change

- Fixed stateless executor out-of-context handling
 - Fixed token tests

- Added ifdef to remove `Deconstruct` extension on everything except `NETSTANDARD2_0`

- Covered text transforms
 - Removed unnecessary non-async transforms

- Replaced all binaries

- params contains a list of loras, instead of just one

- Removed unused `AddRangeSpan` extension

* Added the low level bindings to beam search.

- Minimal amount of characters converted
   - Allocation free
 - Added `TokensToSpan` to `SafeLlamaModelHandle` which converts as many tokens as possible into a character span
   - Allocation free

Add GGUF support

- Made all the mechanics of grammar parsing (GBNFGrammarParser, ParseState) internal. Just call `Grammar.Parse("whatever")`.
 - Added a `GrammarRule` class which validates elements on construction (this allows constructing grammar without parsing GBNF).
   - It should be impossible for a `GrammarRule` to represent an invalid rule.

This commit includes changes to compile en VS Mac + changest to use llama2 not codellama.

It includes MacOS binaries in memory and metal

SymbolIds is now SortedDictionary (although I'm not sure it really needs to be) because the test was failing due to expected value being in another order. The C++ data structure if SymbolIds is std::map<std::string, uint32_t> so the items are ordered by key.

Based on this version: 6b73ef1201

ModelsParams record class

- Used those methods to add a `Clone` method to SafeLLamaContextHandle
 - Simplified `LLamaContext` by using the new methods
 - Sealed `LLamaContext` and `LLamaEmbedder`

- Equality, hashing etc all implemented automatically
 - Default values are defined in just one place (the properties) instead of the constructor as well
 - Added test to ensure that serialization works properly

- Deleted `NativeInfo` (internal class, not used anywhere)