Various improvements

[saharNooby]

This time, there are actually new features and QoL improvements for end-users!

• added rwkv_eval_sequence_in_chunks: an easy to use function for processing whole prompts, instead of splitting them in chunks manually (which is prone to errors) or using eval (which is very slow)
• added model head offloading: helps to cut ~10 ms on my machine when using CUDA
• removed dependency on PyTorch for inference in Python; you still need PyTorch for model conversion and LoRA application
• removed dependency on tokenizers for World models inference in Python; you still need tokenizers to run Pile and Raven models
• added function gpu_offload_layers to RWKVModel, now you don't need to guess how many layers you need to offload before creating the model: you can create the model, get n_layer and call gpu_offload_layers after that
• tokenizer argument is now optional in Python scripts, it will be guessed from n_vocab of the loaded model

Closes #106