Fix mask dtype mismatching query dtype properly (#1984)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/1984

Some refactoring on the mask handling in llama.

In our quantized group embedding implementation we directly call llama_quantized.embedding_type but that operator always return fp32. We need to convert it back to fp16 based on the dtype we got in the QuantizedGroupEmbedding module.

Reviewed By: kimishpatel

Differential Revision: D53796343

fbshipit-source-id: 09537724bc648a1e2bf00bb9be2c56b19894acd1