fix: replace bare except with extension-based dispatch in TorchScript model loading

[AdityaX18]

Body:
Fixes #419

## Bugs Fixed (3 total across 2 files)

| File | Class | Bug |
|---|---|---|
| `torch_detection.py` | `TorchImageDetectionModel` | `except Exception` catches warnings, prints false message |
| `torch_segmentation.py` | `TorchImageSegmentationModel` | bare `except:` catches everything + missing `weights_only=False` |
| `torch_segmentation.py` | `TorchLiDARSegmentationModel` | `except Exception` + missing `weights_only=False` |

## Root Cause
All three used try/except to detect model format. `torch.jit.load()` on 
a valid `.torchscript` file can raise an internal `UserWarning` before 
succeeding. The bare except catches this, prints a false diagnostic, and 
loads correctly via fallback anyway — misleading the user.

The bare `except:` in `TorchImageSegmentationModel` is additionally unsafe: 
it catches `KeyboardInterrupt`, `SystemExit`, and `GeneratorExit`.

## Fix
Primary dispatch by file extension. Fallback probe only for unknown 
extensions, catching only `RuntimeError`, emitting `warnings.warn()` 
(not `print()`) with actionable guidance.

## Tests

pytest tests/test_torch_model_loading.py -v
4 passed

> Note: I am a GSoC 2026 applicant for JdeRobot. This fix directly 
> supports robust model loading as a prerequisite for the robustness 
> evaluation module proposed in my application.Fixes #419

Three bugs fixed across two files:

1. torch_detection.py: bare 'except Exception' caused valid .torchscript files to trigger the except block on an internal PyTorch UserWarning, printing 'Model is not a TorchScript model' even for valid TorchScript.

2. torch_segmentation.py TorchImageSegmentationModel: bare 'except:' with no exception type — catches KeyboardInterrupt, SystemExit, everything. Also missing weights_only=False on torch.load().

3. torch_segmentation.py TorchLiDARSegmentationModel: same bare except Exception pattern + missing weights_only=False.

Fix: primary dispatch by file extension (.torchscript -> torch.jit.load, .pt/.pth -> torch.load with weights_only=False). Unknown extensions probe TorchScript first, catch only RuntimeError, fall back to torch.load() and emit warnings.warn() with actionable guidance instead of print().

Added tests/test_torch_model_loading.py covering all 4 loading paths.

[dpascualhe]

Closing since #419 has already been resolved. Thaks for your interest!