Make TorchTune Llama model KV cache compatible in eager

[jackzhxng]

Summary

• Set up KV caches for TorchTune Llama model
• Adds a separate runner for TorchTune Llama models, since the input handling methods are separate enough to warrant a new copy

Test plan

Download checkpoint: tune download meta-llama/Llama-3.2-11B-Vision-Instruct --output-dir /tmp/Llama-3.2-11B-Vision-Instruct

Eager without KV cache:

python -m examples.models.llama3_2_vision.runner.eager --model llama3_2_vision --checkpoint /tmp/Llama-3.2-11B-Vision-Instruct/original/consolidated.pth  --params examples/models/llama3_2_vision/text_decoder/params/demo_config.json --metadata '{"append_eos_to_prompt": 0, "get_bos_id":128000, "get_eos_ids":[128009, 128001], "get_n_bos": 0, "get_n_eos": 0}' --output_name="llama3_2_vision.pte" --tokenizer /tmp/Llama-3.2-11B-Vision-Instruct/original/tokenizer.model -d fp32 --verbose --prompt "What is the capital of USA?" --max_seq_length 64

Eager with KV cache:

python -m examples.models.llama3_2_vision.runner.eager --model llama3_2_vision --checkpoint /tmp/Llama-3.2-11B-Vision-Instruct/original/consolidated.pth  --params examples/models/llama3_2_vision/text_decoder/params/demo_config.json --metadata '{"append_eos_to_prompt": 0, "get_bos_id":128000, "get_eos_ids":[128009, 128001], "get_n_bos": 0, "get_n_eos": 0}' --output_name="llama3_2_vision.pte" --tokenizer /tmp/Llama-3.2-11B-Vision-Instruct/original/tokenizer.model -d fp32 --verbose --prompt "What is the capital of USA?" --max_seq_length 64 -kv

[larryliu0820]

Please add test in a follow up