[1/n] Lightweight Ray AIR API refactor (ray-project#36706)

This PR removes some circularities in the Ray AIR import system so we can put the training related functions into `ray.train`. It introduces a training context and makes report, get_dataset_shard, Checkpoint, Result, and the following configs:

- CheckpointConfig
- DataConfig
- FailureConfig
- RunConfig
- ScalingConfig

available in `ray.train`. No user facing changes yet, the old APIs still work.

Going forward, it will be most consistent / symmetrical if these things are included in the following way:

```python
from ray import train, tune, serve # Pick the subset that is needed
from ray.train import CheckpointConfig, DataConfig, FailureConfig, RunConfig, ScalingConfig

def train_func():
    dataset_shard = train.get_dataset_shard("train")
    world_size = train.get_context().get_world_size()
    # ...
    train.report(...)

trainer = train.torch.TorchTrainer(
    train_func,
    scaling_config=ScalingConfig(num_workers=2),
)
result = trainer.fit()
```

We have many examples in ray-project#37123 on how this looks like in actual code.

python/ray/train/__init__.py

@@ -2,8 +2,10 @@
 from ray.train._internal.data_config import DataConfig
 from ray.train._internal.session import get_dataset_shard, report
 from ray.train.backend import BackendConfig
-from ray.train.constants import TRAIN_DATASET_KEY
+from ray.train.data_config import DataConfig
 from ray.train.context import get_context
+from ray.train.constants import TRAIN_DATASET_KEY
+from ray.train._internal.session import get_dataset_shard, report
 from ray.train.trainer import TrainingIterator
 
 from ray.air import Checkpoint