Update model

[dbaranchuk]

• Default for client's model parameters: requires_grad=False;
• Interface: "abstract" class BloomForYou. DistributedBloomForYou inherits from it. All consequent classes for various Bloom applications inherit DistributedBloomForYou;
• lm_head -> h @ word_embs for DistributedBloomForCausalLM;
• Updated converted_model without lm_head: dbaranchuk/test-bloomd-6b3;
• h @ word_embs is very slow in fp16 on CPU. Let's keep fp32 for now. TODO @dbaranchuk: understand why fp16 is slow on CPU. If everything is correct then (TODO @dbaranchuk) keep embeddings in fp16 but calculate fp32 casting on the fly;
• Minor bug fix in run_local_server.sh;
• Updated tests/test_full_model.py:
  Test results for REF_NAME='bigscience/bloom-6b3':
  CPU vs GPU (А100) (fp32) passes with (forward_atol=1e-2, inference_atol=1e-2)
  CPU vs CPU (fp32) passes with (forward_atol=1e-4, inference_atol=1e-3)