Support transformer-style tu- encoders in UnetPlusPlus and Linknet

[ShunsukeKikuchi]

Hi @qubvel,

Fixes #1260.

This PR fixes decoder compatibility for transformer-style tu- encoders in UnetPlusPlus and Linknet.

Problem

TimmUniversalEncoder inserts a zero-channel placeholder for the missing 1/2-scale feature in transformer-style backbones such as ConvNeXt / ConvNeXtV2 (tu-convnext_atto, tu-convnextv2_base, etc.).

This works for some decoders, but UnetPlusPlus and Linknet currently fail when they encounter that zero-channel stage.

Changes

UnetPlusPlus

• Treat zero-channel dense decoder nodes as explicit no-op tensors
• Keep the existing TimmUniversalEncoder behavior unchanged

Linknet

• Ignore zero-channel skip connections during skip fusion
• Keep decoder stream channels non-zero when the encoder exposes a zero-channel placeholder stage

These changes make decoders explicitly handle the missing 1/2-scale stage rather than trying to materialize invalid zero-channel operations.

Tests

Added regression tests for tu-convnext_atto on both UnetPlusPlus and Linknet

1. Ran regression tests for the affected models and timm universal encoders:

uv run --with-requirements requirements/test.txt pytest \
  tests/models/test_unetplusplus.py \
  tests/models/test_linknet.py \
  tests/encoders/test_timm_universal.py -q

Result: 30 passed, 2 skipped, 36 subtests passed

2. Ran the repository standard non-marked suite:

uv run --with-requirements requirements/test.txt pytest -v -rsx -n 2 tests/ --non-marked-only
Result:

Result: 198 passed, 7 skipped, 213 subtests passed

[qubvel]

Looks good, thanks for fixing!