[Test][Refactor] Decouple test_dataloader according to the accelerator relevance. (#189250)

## Summary
- Refactored `test/test_dataloader.py` following the S1/S2/S3 decoupling strategy:

### test_dataloader.py
- **S1 (accelerator-unrelated)**: `TestDataLoader` — 66 tests using no device APIs.
- **S2 (accelerator-agnostic)**: `TestDataLoaderDevice` (renamed from
  `TestDataLoaderDeviceType`) — pin_memory tests using `instantiate_device_type_tests`.
  `TestDictDataLoaderDevice` — pin_memory tests with accelerator-generic guards.
- **S3 (accelerator-specific)**: `TestDataLoaderCUDA` and `TestDataLoaderCUDAPersistentWorkers`
- Renamed `TestDataLoaderDeviceType` → `TestDataLoaderDevice` per S2 convention.

### No test logic changed
Only tests were split/moved between classes. No new tests added, none removed.

## Test evidence
Test count preserved:
```
test/test_dataloader.py:     126 → 126 ✓
```

Evidence commands:
```bash
$ grep -c 'def test_' test/test_dataloader.py
126

```

## Refactoring strategy
Per the refactor-test-decoupling methodology:
- S3 > S2 > S1 classification hierarchy strictly followed.
- Whitelist decorators (`@onlyCUDA`, `@unittest.skipIf(not TEST_CUDA, ...)`)
  enlarged to `@onlyAccelerator` / S3 class extraction.
- Blacklist decorators (`@skipIfXpu`, `@skipIfRocm`, `@onlyNativeDeviceTypesAnd`)
  preserved as-is.
- Category C APIs (CUDA IPC) are the only patterns that make a test S3.
- `instantiate_device_type_tests` used for S2 classes.

Authored with an AI assistant.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/189250
Approved by: https://github.com/guangyey, https://github.com/fffrog, https://github.com/Niran814804102, https://github.com/jansel