feat: add range search across language bindings

[JunRuiLee]

Summary

Expose the core range-search functionality from #108 through the existing C ABI/generated header, C++, JNI/Java, and Python bindings.

• Support IVF-Flat, IVF-SQ, IVF-PQ, and IVF-RQ with L2, cosine, and inner product, for single/batch queries with or without a serialized Roaring allow-list.
• Delegate raw half-open distance bands and public endpoint conversion to core, including structural unboundedness, strict/inclusive boundaries, and inner-product direction reversal.
• Return variable-length CSR results with per-query statistics, call-level list reads, and reader capability queries.
• Add explicit native-result ownership, exception-safe cleanup, checked lengths/conversions, and validation/error propagation.
• Preserve existing top-K APIs and behavior. Python still imports and executes top-K against older/partial native libraries; unavailable range APIs report a clear upgrade error.

Tests and documentation

• Add a core-only oracle generator and 168 shared cases consumed independently by C, C++, Java, and Python. Cases cover the full family/metric/API matrix, bounded/unbounded/empty bands, and absent/non-empty/empty filters.
• Compare per-query label/distance-bit multisets, CSR offsets, and all counters without imposing an ordering guarantee absent from core.
• Add boundary, overflow, malformed-input, ownership, callback/reentry, old-library compatibility, and unaligned-NumPy-input regression tests.
• Document each language API and ownership contract; wire oracle generation and consumption into all four binding CI jobs.

Validation

Validated locally on macOS:

• cargo fmt --all -- --check
• cargo clippy --all-targets --workspace -- -D warnings
• cargo test --workspace: 676 passed, 2 ignored
• C11/C++17 strict builds (-Wall -Wextra -Werror), existing and new binding tests, and ASan/UBSan runs: passed; 168 oracle cases per binding
• Java Maven API tests, native validation/panic-boundary/handle-safety tests, and the range oracle under -Xcheck:jni: passed; 168 oracle cases
• Python tests with PVI_RANGE_FIXTURES enabled: 497 passed, including 168 oracle cases
• ASF license-header checks and git diff --check: passed

Independent pre-PR reviews were completed; the older-native-library and unaligned-input findings were fixed and re-reviewed. C++/JNI allocation-failure paths were reviewed but are not exhaustively fault-injected. Cross-platform CI remains to run on this PR.

Scope and compatibility

Based on main at 0e25175 (#108). No core source/tests, storage-format, algorithm, or performance-tuning changes are included. DiskANN remains unsupported for range search. No data migration is required; rollback is a revert of this PR.

[jerry-024]

[Major] Make raw distance semantics explicit across bindings

fromEndpoints accepts public-space values, but range results expose core values as plain distances (L2 is squared and inner product is negated). A caller that compares a returned value with the original endpoint therefore gets the wrong answer. Please apply one explicit contract across Rust, C, C++, Java, and Python: name result values rawDistances/raw_distances (including indexed accessors), and make raw-band constructors internal or explicitly raw-named so the public construction path is fromEndpoints.

[JunRuiLee]

Thanks, fixed in 3cdd045. Unified raw_distances/rawDistances (including indexed/query access) across all five languages, with explicit raw constructors and public endpoint factories. Added API-contract/endpoint regressions and updated docs; numeric values and zero-copy behavior are unchanged.

[jerry-024]

+1