Improve bfloat16 serialization (backward compatible)

[justheuristic]

• added bfloat16 serialization that sends 2 bytes per value (previously, we sent 4);
• changed de-serialization code so it supports both modes of serialization.
• the new mode can be enabled via export LEGACY_BFLOAT16=0
• tested in pytorch 1.12 and 1.13

Next minor release: switch to new serialization mode
Next major release: remove legacy mode

[codecov]

Codecov Report

Merging #553 (90527a7) into master (39f6f81) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master     #553   +/-   ##
=======================================
  Coverage   75.93%   75.93%           
=======================================
  Files          81       81           
  Lines        7982     7995   +13     
=======================================
+ Hits         6061     6071   +10     
- Misses       1921     1924    +3     

Impacted Files	Coverage Δ
hivemind/compression/base.py	94.44% <100.00%> (+1.22%)	⬆️
hivemind/moe/server/connection_handler.py	46.87% <0.00%> (-1.05%)	⬇️
hivemind/moe/server/runtime.py	69.16% <0.00%> (-0.84%)	⬇️
hivemind/moe/server/server.py	43.71% <0.00%> (-0.55%)	⬇️

[borzunov]

I would suggest to:

• Rename LEGACY_BFLOAT16 to USE_LEGACY_BFLOAT16 (both the env var and the constant), so that it's clear that it is a bool flag.
• Remove the LEGACY_PYTORCH constant and use hasattr() instead, so we do not depend on knowing the exact version where these fields were renamed.