Fix torch import guards in rtdetr_pose to allow execution without torch installed

[Copilot]

Several rtdetr_pose modules imported directly from torch at module level with no fallback, causing ModuleNotFoundError in any environment without torch — even when the code path being exercised has no torch dependency (e.g. argument parsing, run-dir defaulting).

Root cause

The import chain tools/train_minimal.py → rtdetr_pose.train_minimal → factory → models.backbones hit these unguarded imports before any torch-optional logic could execute:

• backbone_interface.py — from torch import Tensor, nn
• backbone_projector.py — from torch import Tensor, nn
• models/backbones/blocks.py — import torch; from torch import nn
• models/backbones/csp_backbones.py — from torch import Tensor, nn
• models/backbones/torchvision_backbones.py — from torch import Tensor

Changes

• Wrapped all five unguarded torch imports in try/except ImportError stubs, matching the existing pattern in model.py, losses.py, factory.py, etc.
• backbone_interface.py uses a named stub _NNModuleStub rather than object as the nn.Module fallback — class BaseBackbone(object, ABC) triggers a Python MRO error; a non-object-rooted stub resolves it cleanly:

# backbone_interface.py
try:
    from torch import Tensor, nn
except ImportError:
    class _NNModuleStub:
        pass
    Tensor = None
    nn = SimpleNamespace(Module=_NNModuleStub)

class BaseBackbone(nn.Module, ABC):  # MRO valid with stub, broken with object
    ...

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.