fix(prune): PruneCallback logs the applied ratio and traces on the fit's own batch

[nathanhubens]

Two PruneCallback fixes, follow-up of #98.

1. The epoch log printed the previous step's ratio

torch-pruning's step() increments current_step then applies per_step_pruning_ratio[current_step]. After completed_steps steps, the applied level is therefore levels[completed_steps], but after_epoch printed levels[completed_steps - 1]. The last epoch never showed the final target ratio.

• Now: levels[completed_steps], with the guard completed_steps < len(levels).
• RED on master, from the new test that cross-checks the printed ratio against torch-pruning's applied level: printed 4.05 % / 41.43 % / 49.87 % against 21.06 % / 46.38 % / 50.00 % applied. GREEN: the last epoch prints 50.00 %.

2. The Pruner traced a fixed 224×224 input

before_fit read dls.one_batch() but never passed it on, so the Pruner traced its default torch.randn(1, 3, 224, 224) instead.

• Now the fit's own batch is passed as example_inputs. An example_inputs given by the user still wins.
• RED on master: a model whose head is sized for 8×8 inputs failed in before_fit (mat1 and mat2 shapes cannot be multiplied). GREEN: it traces on the (16, 3, 8, 8) batch, on the model's device, and a user-supplied (2, 3, 8, 8) input is kept.

Tutorial

tutorials/prune/prune_callback was re-executed, since its stored epoch-log lines change. It is a single run with no seed, so every printed number moved. The prose was updated to the printed values and checked by a results review:

• 4800 → 3349 channels (30.2 %);
• MACs ratios 0.6335 / 0.6346;
• parameter ratio 0.18, unchanged.

The review also caught a pre-existing error: the head goes from 512 to 502 units, not from 1024. 1024 is the layer's input width.

Evidence

• Full suite nbdev-test passes, as do the slow prune tests.
• Clean and export leave the tree unchanged with CI's nbdev 3.3.21.