update datasets dependency to >=3.0.2,<4.0.0 (results in 3.6.0)

[ArneBinder]

We require >=3.0.2 because of huggingface/datasets#7234, otherwise Brat datasets break (since datasets>=2.16.0).

related: #93, #212

requires:

• implement test_base_builder for brat #215
• Fix PieDatasetBuilder #216
• fix builder fixture #217
• make empty_dataset fixtures future proof #218

dataset scripts require:

• add missing user to BASE_DATASET_PATHs #219

Changes:

• use trust_remote_code=True in tests (required since datasets 2.20.0)
  • add hack to test_load_dataset_conll2003(_single_split): pass base_dataset_kwargs=dict(trust_remote_code=True). This should not be necessary anymore once this is handled in the conll2003 pie dataset script
• we set trust_remote_code=True per default pie_datasets.load_dataset and in pie_datasets.DatasetDict.load_dataset to stay backwards compatible
• adjust some builder tests since exception types and messages did change: test_builder_class_with_kwargs_wrong_parameter and test_builder_class_with_base_dataset_kwargs_wrong_parameter
• also test DatasetDict.load_dataset in test_load_dataset_conll2003
• remove dependency restrictions, they shouldn't be required anymore: numpy = "<2.0.0", pyarrow = "^13", and fsspec = "<2023.9.0"

Upgrade instructions: Adjust PIE dataset scripts to work with trust_remote_code if their base dataset requires it. Either

1. add entries {"trust_remote_code": True} for all config names and None to BASE_CONFIG_KWARGS_DICT, or
2. switch to a parquet based base dataset (there is often a branch on the HF hub with the converted dataset). However, this may required to derive the dataset builder class from pie_datasets.ArrowBasedBuilder (instead of pie_datasets.GeneratorBasedBuilder).

TODO:

• test in downstream project (should cover at least conll2002, drugprot and sciarg), see use pie-datasets from branch update-datasets-dependency pytorch-ie-hydra-template-1#219

Follow-ups:

• update all dependencies to latest version via poetry update
• update dataset scripts

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.62%. Comparing base (f687772) to head (3e42d27).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #213      +/-   ##
==========================================
- Coverage   97.47%   93.62%   -3.85%     
==========================================
  Files           5       10       +5     
  Lines         396      989     +593     
==========================================
+ Hits          386      926     +540     
- Misses         10       63      +53     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.