feat(regularize): GroupRegularizeCallback for sparse training before a structured prune

[nathanhubens]

Adds GroupRegularizeCallback, the training-loop companion of Pruner.regularize (#96). It implements the sparse-training protocol of Network Slimming / DepGraph: regularize during a fit, prune once after it, then fine-tune.

pruner = Pruner(learn.model, 0.3, 'local', large_final, reg=1e-4, example_inputs=xb)
learn.fit(10, cbs=GroupRegularizeCallback(pruner, schedule=lin))   # group penalty on the gradients
pruner.prune_model()                                                 # once, after the fit
learn.fit(2, reset_opt=True)                                         # the prune replaced Parameters: rebuild the optimizer

Mechanism

• Order MixedPrecision.order - 2:
  • after GradientAccumulation, which cancels the micro-batches, so the penalty is applied once per real optimizer step;
  • before MixedPrecision unscales the gradients, so the penalty is multiplied by the loss scale of a real GradScaler (1 otherwise).
• Optional Schedule ramps the penalty. The callback keeps its own copy, so the module-level schedule objects other callbacks rely on are never mutated.
• Refused, each with a one-line error:
  • a Pruner built with reg == 0;
  • a Pruner built on another model than learn.model;
  • NonNativeMixedPrecision;
  • a PruneCallback in the same fit (mid-fit pruning would invalidate the penalized groups).
• PruneCallback now refuses reg: before, it was forwarded to the Pruner without ever being applied.
• Lives in regularize/, next to RegularizeCallback. The RegularizeCallback BatchNorm caveat and the Pruner "Group sparse learning" section both point to it.

Evidence

• nbs/tests/test_group_regularize_callback.ipynb:
  • hand-computed gradients through a real fit step;
  • a loss-scale stub on CPU (the gradients are scaled, then unscaled);
  • GradientAccumulation applies the penalty once per step;
  • schedule ramps, with the shared schedule left untouched;
  • every refusal;
  • fit, then prune_model(), then fit(reset_opt=True): the optimizer holds exactly the live parameters, while without reset_opt it holds stale ones;
  • the order pinned between GradientAccumulation and MixedPrecision;
  • a #| slow ResNet-18 run ending in learn.save.
• Three mutations were run: wrong order, loss scale not read, schedule not copied. Each one fails its test.
• Full suite nbdev-test and the slow tests pass in a clean clone with CI's nbdev 3.3.21. Clean and export leave the tree unchanged. No circular import between prune and regularize.
• No measured claim: whether sparse training improves accuracy after pruning is not evaluated here.