Do not identify FLUX ControlNet checkpoints as FLUX main models

[chiliec]

Summary

Installing an XLabs FLUX ControlNet checkpoint (e.g. flux-canny-controlnet-v3.safetensors from
XLabs-AI/flux-controlnet-collections) registered it as a FLUX main model (type: main,
base: flux, format: checkpoint) instead of a ControlNet, so it never appeared where ControlNets
are chosen and could not be used.

The XLabs ControlNet contains two FLUX double blocks, so its state dict carries
double_blocks.0.img_attn.norm.key_norm.scale — the key Main_Checkpoint_FLUX_Config._validate_is_flux
keys off. ControlNet_Checkpoint_FLUX_Config matches too, and because
ModelConfigFactory.matches_sort_key ranks ModelType.Main first, the main config won the tie-break.

_validate_is_flux now rejects state dicts that is_state_dict_xlabs_controlnet or
is_state_dict_instantx_controlnet recognise, so only the ControlNet config matches these files.

Related Issues / Discussions

Closes #9832

QA Instructions

New regression test tests/backend/model_manager/configs/test_flux_controlnet_not_main_config.py
builds a state dict mirroring the XLabs ControlNet (the controlnet-specific keys plus two FLUX double
blocks, no single_blocks/final_layer) and asserts the main config rejects it while the ControlNet
config accepts it.

python -m pytest tests/backend/model_manager/configs/test_flux_controlnet_not_main_config.py

• Before the fix: test_xlabs_controlnet_does_not_match_flux_main fails (DID NOT RAISE NotAMatchError —
  the file matched as a main model).
• After the fix: both tests pass (2 passed).

Verified locally on a CPU-only Linux box (Python 3.12, torch 2.13.0+cpu). The FLUX/ControlNet/main
config tests pass (67 passed, 1 skipped); the remaining failures in the wider
tests/backend/model_manager/configs/ directory are pre-existing on a clean main in this
environment (protobuf/onnx and WAN filename-heuristic env issues) — the same count fails with and
without this change.

Review

Smallest fix at the root cause: two is_state_dict_*_controlnet detectors already exist and are used
by the ControlNet config; reusing them in the main-model guard keeps behavior consistent and avoids
touching the fragile ControlNet key heuristics.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[Pfannkuchensack]

Thanks! Verified with the real flux-canny-controlnet-v3.safetensors: on main both ControlNet_Checkpoint_FLUX_Config and Main_Checkpoint_FLUX_Config match and main wins; with this PR only the ControlNet config matches. Real FLUX.1 checkpoints (flux1-dev-fp8, raw and scaled) are still identified as FLUX main models, and the new test fails without the fix.

One change before merging: the is_state_dict_instantx_controlnet check in _validate_is_flux can't be reached. InstantX ControlNets use diffusers keys (transformer_blocks.*, controlnet_x_embedder.*) and have no double_blocks.0.img_attn.norm.key_norm.scale, so the key check just above already rejects them. Please drop it and keep only the XLabs check, unless you have a real InstantX file with BFL double_blocks keys.

Otherwise this looks good to me.

[Pfannkuchensack]

Need to clean up the conflict but nice work.