Use dependency injection for runner (#53)

Summary:
X-link: pytorch/executorch#10326


Pass in runner components, move most of the instantiation logic from `load()` to a new static API `create()`.

This adds testability to runner components.

Next step would be moving most of the logic out into `extension/llm/runner/` so that it can be used on non-llama models.

Currently the logic for getting tokenizer instance should not assume llama, which I can modify in next diff.

Differential Revision: D73165546

Author: larryliu0820

include/pytorch/tokenizers/tokenizer.h

@@ -56,6 +56,10 @@ class Tokenizer {
     return eos_tok_;
   }
 
+  virtual bool is_initialized() const {
+    return initialized_;
+  }
+
  protected:
   bool initialized_ = false;
   int32_t vocab_size_ = 0;