string dtype fixes

[d-v-b]

this PR fixes the issues reported in #3167.

There were two problems driving that issue:

• zarr python was not using the string "str" to denote a variable length string data type. Instead, in main we use NumPy's behavior and map "str" to a fixed-length UTF-32 string, which is not a useful data type for most people.

  In this PR, when a user requests dtype=str or dtype="str" or dtype="string" they get a VariableLengthUTF8 dtype instead of the nearly-useless U dtype.

• zarr python could create data types for length-0 scalars. Since these data types cannot be associated with arrays that contain any values, they are useless. See disallow 0-length fixed-size data types #3168. This is a quirk of NumPy that we don't need to support. This PR contains logic to ensure that we don't create data types for length-0 scalars.

note that while debugging this, I found that zarr-python 2.x could create arrays with where the dtype in metadata indicates length 0, but the actual array uses a dtype with size 1. We will need some support for this, but I don't think we need that in this PR.

[d-v-b]

another important change: I altered the zarr v3 name of the VariableLengthUTF8 dtype to be "string" instead of "variable_length_utf8". This should resolve compatibility issues with zarr-python 3.0.8.

[ianhi]

This line leaves me a little confused on what the intended behavior for FIxedLengthString with an empty fill value is. or in general does the fill vlaue of a fixed length string need to be the same length?

[d-v-b]

I'm going with numpy's semantics, which assigns the empty string to a sequence of null bytes:

>>> np.array([''], dtype="U1").tobytes()
b'\x00\x00\x00\x00'

[d-v-b]

this is a situation where the array data type and the scalar data type (the thing you get when you index that array) are not the same:

>>> np.dtype('U0').type('a').dtype
dtype('<U1')
>>> np.dtype('U0').type('').dtype 
dtype('<U')
>>> np.array([''], dtype="U0").dtype   
dtype('<U1')
>>> np.array([''], dtype="U0")[0].dtype
dtype('<U')

[ianhi]

Good from me! Seems to fix the issue I was seeing with Xarray.

[d-v-b]

for posterity, numpy does permit the creation of arrays with a "V0" data type, i.e. raw bytes where each element has length 0:

>>> np.array([b''], dtype="V0").dtype
dtype('V')
>>> np.array([b''], dtype="V0").tobytes()
b''
>>> np.array([b'',b'',b'',b''], dtype="V0").tobytes()
b''

Arrays with this data type are all the empty byte string b"". I don't think this is meaningful for zarr.

The changes in this PR will make this data type unrepresentable in zarr python. I hope we are OK with this.