Add initial support for RWKV v7

[MollySophia]

Changes:

• Update chat roles of QA mode in python chat example
• Sync latest ggml upstream
• Use the WKV kernel in upstream ggml for v5/v6
• Refactor & simplify v4/v5/v6 functions, offload more possible tensors to gpu (better performance when using gpu)
• Get rid of _inplace operations (ggml scheduler will handle these and use inplace variant automatically when possible)
• RWKV v7 support with initial cpu wkv op (with multithreading)
• q4_k/q5_k quants support

Currently tests done:

• Inference of v7 0.1B & 0.4B with or w/o offloading with CUDA, with or w/o quantization. The outputs aligns with ChatRWKV
• A quick lm_eval of lambada_openai on v7 0.1B gives 12.407, which matches the data from Bo

TODO:

• Add tiny-rwkv-v7 for unit test Done
• Update README.md (Added some info about rwkv v7. quality/perf test data haven't been added yet)
• Fix ROCm build Still cannot get working ROCm build. Maybe fix it next time?
• Do more tests

Tested:

• ArchLinux with CPU backend ✅
• ArchLinux with CUDA 12.6 ✅
• Windows with CPU backend ✅
• Windows with CUDA 12.2 ✅
• macOS, m2 with CPU backend ✅
• macOS, m2 with metal backend ✅

[harrisonvanderbyl]

Merged, can have rocm and and tests in separate PR

[MollySophia]

Merged, can have rocm and and tests in separate PR

Unit tests for v7 are already there too. There's only ROCm left