Modelly 1.4.2

A patch release for numpy datasets (DataRegistry.track_agents_numpy): their fields now belong
to each model, so two models alive at once, a model and its copy from Model.load, and a
checkpoint resumed in a new process all read the right values. No core operation is slower than
1.4.1 (python -m modelly.dev perf).

Fixed

• The fields of a numpy dataset were properties of the agent class bound to one dataset's array,
  the last one made. Two live models of the same agent class read and wrote the second model's
  values; a model restored by Model.load read the original model's array; a checkpoint loaded
  in a fresh process had no fields at all (AttributeError). Each agent now holds a reference to
  its own dataset, a restored dataset puts the fields back on the class, and checkpoints written
  by 1.4.0 and 1.4.1 load and resume as before. Closing one model's registry leaves the fields
  in place while another model's dataset of that name is open. Reading or writing a tracked
  field costs about 20 ns more: 1.07x on a step that only updates two tracked fields of 10,000
  agents (seven paired rounds on one core).
• SQLDataRecorder stores a numpy dataset's agent_id as an integer, as the memory and Parquet
  recorders do, instead of a float.