Fix feature_info channels for pruned models

[prawnsgupta]

Summary

Pruned models report incorrect feature_info.channels() — the original (unpruned) channel counts rather than the actual pruned feature-map channels. This fixes #2370.

Affected models: efficientnet_b1_pruned, efficientnet_b2_pruned, efficientnet_b3_pruned, ecaresnet50d_pruned, ecaresnet101d_pruned.

Root cause

adapt_model_from_string() in timm/models/_prune.py rebuilds Conv2d/BatchNorm/Linear layers with pruned dimensions, but feature_info is populated at original build time and was never updated. So feature_info.channels() (and anything built on it, e.g. features_only=True) kept the dense channel counts.

Before (efficientnet_b1_pruned, features_only=True):

actual feature shapes : [16, 12, 35, 67, 320]
feature_info.channels : [16, 24, 40, 112, 320]   # wrong

Fix

Adds _adapt_feature_info(), called at the end of adapt_model_from_string(). It runs a single dry-run forward with hooks on the feature modules to read their true output channel counts and writes them back into feature_info. Properties:

• Fixes all pruned models, since it lives in the shared _prune utility rather than any single model file.
• Fail-safe — wrapped so any error leaves feature_info untouched and never breaks model construction (logs a warning).
• Handles both the plain-list and FeatureInfo representations, and derives input channels/size from the model so it respects custom stems and pretrained_cfg input size.

After the fix, feature_info.channels() matches the real feature shapes for all five models.

Tests

Pruned models now pass the feature-extraction tests, so the *pruned* exclusions in tests/test_models.py (one marked # hopefully fix at some point) are removed. Verified locally:

• test_model_forward_features, test_model_forward_intermediates_features, test_model_forward_intermediates — 28 passed for pruned models (previously excluded).
• Non-pruned feature tests unaffected.

Note: the separate in_chans-on-pruned limitation (#1597) is out of scope here and left untouched.