[python] Avoid Bloom false negatives for floating-point BTree keys

[TheR1sing3un]

Purpose

Python's floating-point BTree comparator treats different NaN payloads as equal, and likewise treats positive and negative zero as equal. The optional Bloom filter hashes serialized bytes, so equality/IN queries using an equivalent value with different bytes can incorrectly return no rows.

Route NaN and zero point queries through the existing comparator-based range lookup. This preserves current comparison semantics and supports already-written indexes without changing key encodings or the file format. Other point queries retain the Bloom fast path.

Tests

• Cover FLOAT and DOUBLE NaN payloads/signs and both signed-zero directions, with Bloom enabled/disabled, across multiple data blocks, including multiple rows sharing a key and equality/IN queries.
• btree_bloom_filter_test.py: 12 passed, including the existing assertion that missing ordinary keys read only the Bloom filter. Seven new cases fail against the unchanged base; one signed-zero case happens to be a Bloom false positive.
• Existing btree_thread_safety_test.py and global_index_build_test.py: 47 passed.
• Flake8 for changed files and git diff --check pass.