Fix UTF generation for numpy in property-based tests

[moradology]

This PR fixes the numpy_arrays() hypothesis strategy for Unicode string (e.g. dtypes (<U#>)) by:

1. Ensuring correct string length enforcement using dtype.itemsize // 4 (instead of parsing str(dtype)[2:]).
2. Removing invalid surrogate Unicode characters (U+D800–U+DFFF) to prevent encoding errors.
3. Introduced safe_unicode_for_dtype(dtype), which:

• Computes the correct string length limit for NumPy's UTF-32-based storage.
• Generates only valid UTF-8 text while filtering out surrogate characters.

Fixes #2732

TODO:

• Add unit tests and/or doctests in docstrings
• Add docstrings and API docs for any new/modified user-facing classes and functions
• New/modified features documented in docs/user-guide/*.rst
• Changes documented as a new file in changes/
• GitHub Actions have all passed
• Test coverage is 100% (Codecov passes)

Why modify the default string generation from hypothesis?

• NumPy's dtype='<U#>' uses UTF-32 storage, which can store all Unicode code points, including surrogates (U+D800–U+DFFF).

• UTF-32 does not enforce the UTF-8 restriction that surrogates must be part of a valid surrogate pair.

• This means NumPy can store invalid Unicode sequences without immediate errors:

  import numpy as np
  arr = np.array(["\ud910"], dtype="<U6")  # Allowed by NumPy
  print(arr)  # ['\ud910']

• Surrogate characters are not permitted in UTF-8.

• If we store surrogates in NumPy and later export the text as UTF-8, it results in a UnicodeEncodeError:

  arr[0].encode("utf-8")  # UnicodeEncodeError

[dcherian]

Thanks @moradology !