Column of arrays (2D+) are using unreasonably high memory

## Describe the bug
When trying to store `Array2D, Array3D, etc` as column values in a dataset, accessing that column (or creating depending on how you create it, see code below) will cause more than 10 fold of memory usage.

## Steps to reproduce the bug
```python
from datasets import Dataset, Features, Array2D, Array3D
import numpy as np

column_name = "a"
array_shape = (64, 64, 3)

data = np.random.random((10000,) + array_shape)
dataset = Dataset.from_dict({column_name: data}, features=Features({column_name: Array3D(shape=array_shape, dtype="float64")}))
```

the code above will use about 10Gb of RAM while constructing the `dataset` object.

The code below will use roughly the same amount of memory (and time) when trying to actually access the data itself of that column.
```python
from datasets import Dataset
import numpy as np

column_name = "a"
array_shape = (64, 64, 3)

data = np.random.random((10000,) + array_shape)
dataset = Dataset.from_dict({column_name: data})
dataset[column_name]
```

## Expected results
Some memory overhead, but not like as it is now and certainly not an overhead of such runtime that is currently happening.

## Actual results
Enormous memory- and runtime overhead.

## Environment info
- `datasets` version: 2.3.2
- Platform: macOS-12.5.1-arm64-arm-64bit
- Python version: 3.8.13
- PyArrow version: 9.0.0
- Pandas version: 1.4.4