Ensure KV cache is not returned as output tensor during decode phase for Falcon

[schoi-habana]

This PR follows the change for llama in https://github.com/HabanaAI/optimum-habana-fork/pull/154/files
This increased the maximum batch size of BF16 falcon180b inference from 250 to 316

updated command (remove --reuse_cache)
python ../gaudi_spawn.py
--use_deepspeed --world_size 8 run_generation.py
--model_name_or_path /root/data/falcon/falcon-180b/snapshots/d2ea5531862d4fe907280234990e6380d2befd97/
--use_hpu_graphs
--use_kv_cache
--bf16
--batch_size 316
--max_new_tokens 128
--max_input_tokens 128
--limit_hpu_graphs
--n_iterations 3
--trim_logits
--bucket_internal
--bucket_size 128
--prompt "I've always managed to dodge the bullet and avoid the addictive pull of Pokemon. Leave it to a button-mashing brawler with plastic figurine accessories to finally get me hooked. At first glance, Pokemon Rumble U isn't much to look at. With its simplistic controls and repetitive gameplay, you might feel inclined to dismiss it as yet another cash-in of the popular Nintendo franchise. But despite its faults, there's actually much more to Rumble U than meets the eye, making this a satisfying and"

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[astachowiczhabana]

huggingface#993