Fix bfloat16 serialization for tensors with zero elements

[borzunov]

Follow-up to #553.

Dummy tensors are widely used across Petals, but we get this with hivemind==1.1.6:

Mar 28 23:05:40.339 [WARN] [hivemind.p2p.p2p_daemon._process_stream:440] Handler failed with the exception:                                                                                           
Traceback (most recent call last):                                                                                                                                                     
  File "/home/jheuristic/anaconda3/envs/py38_petals_borzunov/lib/python3.8/site-packages/hivemind/p2p/p2p_daemon.py", line 431, in _process_stream                                       
    async for response in handler(_read_stream(), context):                                                                                                                               
  File "/home/jheuristic/anaconda3/envs/py38_petals_borzunov/lib/python3.8/site-packages/hivemind/p2p/p2p_daemon.py", line 529, in _stream_handler                                       
    async for item in output:                                                                                                                                                                
  File "/storage/hdd1/jheuristic/exp/decentralized/borzunov/petals/src/petals/server/handler.py", line 137, in rpc_inference                                                                              
    hidden_states, prompts, hypo_ids = map(deserialize_torch_tensor, request.tensors)                                                                                                                    
  File "/home/jheuristic/anaconda3/envs/py38_petals_borzunov/lib/python3.8/site-packages/hivemind/compression/serialization.py", line 47, in deserialize_torch_tensor                                      
    return compression.extract(serialized_tensor).requires_grad_(serialized_tensor.requires_grad)                                                                                                         
  File "/home/jheuristic/anaconda3/envs/py38_petals_borzunov/lib/python3.8/site-packages/hivemind/compression/base.py", line 108, in extract          
    if len(serialized_tensor.buffer) // shape.numel() == 4:  # legacy mode: convert to fp32                                                                             
ZeroDivisionError: integer division or modulo by zero  

This happens for the simplest inference queries.

[codecov]

Codecov Report

Merging #560 (67704d7) into master (98531ce) will increase coverage by 0.04%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #560      +/-   ##
==========================================
+ Coverage   75.88%   75.92%   +0.04%     
==========================================
  Files          81       81              
  Lines        8008     8009       +1     
==========================================
+ Hits         6077     6081       +4     
+ Misses       1931     1928       -3     

Impacted Files	Coverage Δ
hivemind/compression/base.py	94.52% <100.00%> (+0.07%)	⬆️

... and 3 files with indirect coverage changes