Fix CI: add API test deps to requirements.txt, make torch imports lazy

[Copilot]

CI python-test job was failing with pytest: command not found and import errors because requirements.txt only listed data/viz libs, not the API or test dependencies. Additionally, api/inference.py imported torch and instantiated SentinelModel at module level, causing ImportError when the FastAPI app was imported in any environment without PyTorch.

Changes

• requirements.txt — added pytest>=7.0.0, fastapi>=0.100.0, httpx>=0.24.0, slowapi>=0.1.7, pydantic>=2.0.0 with minimum version constraints

• api/inference.py — replaced module-level torch/model initialization with a lazy _get_model() singleton; torch is only imported on the first actual inference call:

  # before: torch imported at module load time
  import torch
  device = torch.device(...)
  model = SentinelModel().to(device)  # fails in CI
  
  # after: deferred until first call
  def _get_model():
      global _device, _model
      if _model is None:
          import torch
          from api.core.model import SentinelModel
          ...
      return _model, _device

• api/routes/inference.py — moved import torch and from transformers import ... from module level into _load_model(); run_inference() retrieves the already-loaded module via sys.modules["torch"] instead of a redundant local import