Check and fix chararray string dimension names

[kmuehlbauer]

• Closes to_netcdf from subsetted Dataset with strings loaded from char array netCDF can sometimes fail #6352, Explicit fixed width 'S1' arrays re-encoded creating extra dimension #2899
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst
• New functions/methods are listed in api.rst

[kmuehlbauer]

This is ready for review.

In case of character arrays when encoding the PR aims to:

0. check encoding["char_dim_name"] for something like string10 -> which gets splitted into string (prefix) and 10 (strlen)
1. if the size of the last dimension of the character array isn't equal to strlen, a warning for possible dimension naming clash is issued and a new dimension name is created with the actual array size
2. if no trailing number is found, we check if encoding["original_shape"] is set, if it doesn't corresponds to the actual string length, a warning for possible dimension naming clash is issued and a new dimension name is created with the actual array size
3. if encoding["char_dim_name"] isn't set, we continue creating new dim names with stringN naming scheme.

At this place each variable is handled separately so we do not have the full picture of all needed character array dimensions. Since we now issue warnings and create fitting dimensions names to prevent breakages there are chances we get false positives (eg in the case of several variables with the same size which have all received the same change in size and have encoding["char_dim_name"] = "somestring" without numbering and also original_shape set). I think that additional warnings in those cases might help users in identifying issues and are acceptable.

[kmuehlbauer]

Pydap test failure in all-but-numba CI run might be real. Need to investigate.

Update: Fails in other PR, too. Pydap backend errors are most likely related to recent changes in numpy (2.2.6 works, 2.3.0 fails)