[Python] Expose dataset read limits and bound prepared-series caching

[ColinLeeo]

Indexed TsFileDataFrame reads currently retain every prepared series until the runtime closes and use a fixed descriptor-cache capacity. Applications cannot set these cache limits through the public constructor.

This adds keyword-only max_prepared_series, descriptor_cache_size, max_open_files, query_workers, and query_parallel_min_rows options for use_index=True. Explicit arguments override the corresponding TSFILE_DATAFRAME_* environment variables, followed by built-in defaults. Options are validated before opening dataset resources and fixed for each runtime; subsets share their parent's configuration.

Prepared-series metadata now uses an LRU cache with a default capacity of 4096. Query leases and dependency tracking protect active handles and shared aligned-time metadata during eviction. Closing the cache waits for preparation, active queries, and native-handle cleanup. Cache capacities of zero disable retention; active queries can temporarily exceed the configured entry limit.

The English and Chinese README sections document the options, defaults, indexed-mode requirement, and per-runtime resource semantics.

Validation

• Built the C++ library and Cython extensions from rc/2.5.0 on macOS arm64 with Python 3.12.
• PYTHONPATH=. python -m pytest tests/test_dataset_read_options.py tests/test_dataset_index.py tests/test_tsfile_dataset.py -q: 193 passed, including 53 new cases for configuration precedence, validation, small/disabled caches, aligned reads, concurrent access, and shutdown races.
• Black 26.3.1 checks and git diff --check passed.