Add streaming full-collection traversal across C++/C/Python (relates to #380).
## API
- C++: `Collection::create_iterator(IteratorOptions)` → `DocIterator` (`next()` returns `Result<Doc::Ptr>`: error / `nullptr` EOF / doc; `close()`, idempotent, also run by the destructor).
- C API: opaque `zvec_doc_iterator_t` + `zvec_iterator_options_t` handles; `zvec_collection_create_iterator` / `zvec_doc_iterator_next` / `zvec_doc_iterator_close`; errors mapped precisely to `zvec_error_code_t`.
- Python: `collection.iter_docs()` returns a `DocIterator` (iterator protocol + context manager; prefer `with collection.iter_docs() as docs:`). The snapshot is taken at call time; the iterator closes itself when exhausted, on `with` exit, or via `close()`, releasing the native slot on every path (including early break and exceptions).
## Snapshot semantics
- `create_iterator()` seals the current writing segment on writable collections (read-only collections scan directly, including their writing segment — no flush, no data loss); the snapshot captures the segment set, a deep copy of the delete bitmap, and the schema. Writes after creation are invisible to the iterator; deletions after creation do not affect it.
- `IteratorOptions.output_fields_` selects forward fields (unknown/duplicate names rejected with an error); `include_vector_` controls vector materialization.
## Concurrency (admission control + active-iterator count)
- Iterators and maintenance operations are mutually exclusive via `maintenance_mtx_`: `create_iterator()` fails fast with `FailedPrecondition` while a maintenance operation (optimize, schema DDL, close, destroy) is running.
- While any iterator is open (active-iterator count under the schema lock): schema DDL (create/drop index, add/alter/drop column), destroy and close return `FailedPrecondition`; the destructor path instead logs and waits for open iterators (it cannot report errors); optimize fails at its start. flush, writes and queries are not affected; Stats/Schema/Options (shared lock) are not affected.
- The collection must outlive its iterators: close every iterator before closing/releasing the collection (documented in the C++/C API headers).
## Implementation notes
- Per-segment readers opened lazily (at most one open at a time); deleted rows filtered by `FilteringReader`, wrapped only when the snapshot's delete bitmap is non-empty (`src/db/index/segment/filtering_reader.*`).
- Each batch is materialized column by column in bounded windows of at most 4096 rows (`kMaxRecordBatchNumRows`): a Parquet scan returns a whole row group per ReadNext (up to ~1M rows), so windows cap doc materialization and keep memory constant. Column indices are resolved and validated once per segment reader; scalars/arrays go through the shared column-level converter; vectors are fetched per field using segment-local row ids (`_zvec_row_id_`, correct for compacted segments). Materialization and reader failures are sticky (the error keeps being returned; no partial docs are ever handed out).
- Shared converters in `src/db/index/common/doc_field_converter.*` also serve the SQL engine (`ConvertVectorDataBufferToDocField` / `ConvertArrowColumnToDocFields`; `SegmentImpl::Fetch` keeps its pre-existing implementation).
- Known limitation: docs without a value for a vector field are accepted by insert (the write path only warns and skips) and are indistinguishable from fetch errors today, so iteration with `include_vector` fails on such docs — documented in code with a TODO; tracked in a separate issue.
## Tests
- C++ iterator_test: 22 tests — basic/empty/deleted/close-then-next, include/exclude vector, output_fields selection + rejection, scalar type mapping, 1000-doc integration, read-only collection, performance (100k docs, constant memory), Parquet large-row-group windowed materialization, and concurrency: snapshot isolation under writes, optimize/DDL/close/destroy rejected while open (recovered after the iterator closes), create-iterator rejected while optimize runs, the destructor waits for open iterators, create-iterator-on-closed-collection rejected, slot released by destructor, multiple iterators.
- C c_api_test: 6 iterator tests (basic/concurrent-semantics incl. destroy rejection + FAILED_PRECONDITION mapping, exclude-vector, output-fields, empty, null-args) inside the 75-test C API suite.
- Python test_iter_docs.py: 11 tests (basic fields/vectors, deletion filtering, output fields, isolation, snapshot-at-call-time, iterator protocol, context manager, early close releases the slot).
Zhuanglin Zheng