pfl 0.5.2

v0.5.2

New features

• PyTorchFederatedDataset takes make_user_dataset_fn, so a user's records
  can become any AbstractDataset rather than only a tensor dataset.

Bug fixes

• Bind each rank to its own GPU under torchrun. get_default_device returned
  a bare cuda, which is device 0 in every process, so a multi-process launch
  put the whole job on one GPU while reporting a world size that said otherwise.
• SummaryMetrics no longer reports quantiles below min_bound, which it did
  for any quantile falling in the first histogram bin, and warns once when the
  metric it summarises is absent instead of silently emitting nothing.
• SimulatedBackend shows each postprocessor the metrics the previous ones
  emitted, so a postprocessor can summarise another's per-user metric.