LM head module

[dbaranchuk]

• BloomForCausalLM == BloomForCausalLM from HF transformers;
• Added LMHead module that doesn't duplicate word embeddings;
• This module also allows for efficient chunked forward when embeddings are in half-precision on CPU. The chunk size can be set in DistributedBloomConfig (default: 10000). Default chunk_size provides reasonable performance and consumes extra ~0.6Gb memory instead of ~7.2Gb for the 176B model;
• Moved 'set_requires_grad' from BloomModel to DistributedBloomModel to make BloomModel == BloomModel from HF transformers.

[justheuristic]

LGTM, let's merge