Numpy Serialization is not Unique for FP8 dtypes

```python
import jax.numpy as jnp
import numpy as np

>>> np.arange(10).astype(ml_dtypes.float8_e4m3b11fnuz).dtype.str
'>> np.arange(10).astype(ml_dtypes.float8_e4m3fn).dtype.str
'>> np.arange(10).astype(ml_dtypes.float8_e4m3).dtype.str
'>> np.asarray(jnp.arange(10).astype(jnp.float8_e3m4)).dtype.str
'

Reproduce the Python examples with the FP8 dtypes, then inspect ml_dtypes/_src/dtypes.cc around line 87 and the handling of kNpyDescrKind. Determine how dtype serialization currently identifies these types and review the backward-compatibility constraint. Done requires an agreed resolution for unique serialization, with its compatibility implications established.

Written by the indexing model from the issue text.