[Feature Request] Universal cache invalidation

### Describe the feature request

From my research only TensorRT has proper cache invalidation, once onnxruntime has been updated the outdated caches are causing onnxruntime to crash.

| EP | Auto-invalidation | Version Detection |
|---|---|---|
| TensorRT (ORT 1.20+) | Yes | Runtime |
| MIGraphX | Partial | Compile-time |
| CoreML | No | None |

For now I use this workaround.

```python
import os
import onnxruntime

cache_path = os.path.join('.caches', onnxruntime.get_version_string())

session = onnxruntime.InferenceSession('model.onnx', providers = [('CoreMLExecutionProvider',
{
'ModelCacheDirectory': cache_path
})])
```

### Describe scenario use case

I recommend a universal solution to handle caches, it should not be EP specific.

No source files, tests, or entry points are identified. Start by comparing the listed execution providers and their current cache behavior, using the Python workaround as the concrete scenario. Done means stale caches are invalidated through a universal solution rather than provider-specific handling.

Written by the indexing model from the issue text.