Review fixes:
- build_meta returns unique_names directly (was built then dropped)
- build_interned_names reduced from 27 lines to 3 (takes &[String])
- name_map borrows &str from index instead of cloning per entry
- Mutation methods documented as raising TypeError
- All user-facing docstrings cleaned: no implementation details,
  no FFI/Rust/interning/refcount language
- Rich .pyi stubs with docstrings for IDE hover

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Performance optimizations addressing PyO3 overhead analysis:

1. Zero-copy parse for bytes input (#6): DocumentOwner enum uses
   PyBackedBytes to borrow directly from Python bytes object's
   internal buffer, avoiding a full memcpy of the XML document.
   str input still copies (Python str -> UTF-8 encoding required).

2. Eliminate String intermediaries (#4): All text-returning methods
   (xpath_text, xpath_string, .text, .tail, .get, .keys, .items,
   itertext, text_content, tostring) now return Py<PyString> built
   directly from &str slices. Skips Rust String allocation that
   PyO3 would then copy again into Python.

3. interned_tag_fast (#3): Hot paths (child_tags, descendant_tags,
   make_element_borrowed, make_elements) now accept &IndexWithMeta
   directly, avoiding redundant borrow_dependent() calls in tight
   loops.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

#1 Iterator pre-caching: ElementIterator pre-builds all (tag_idx,
   cached_tag) pairs in a single Document borrow at creation time.
   __next__ no longer borrows Document — just clone_ref on pre-cached
   values.

#2 Lazy ElementList: Element.xpath() and CompiledXPath.eval() now
   return ElementList — holds one Py<Document> + Vec<usize> of tag
   indices. Elements created on demand via __getitem__/__iter__.
   compiled.eval() for 100K results: 4ms -> 0.07ms (57x faster).
   Supports __len__, __getitem__, __iter__, __bool__, __eq__ (with
   list comparison).

#7 O(1) sibling lookup: child_positions[i] stored in IndexWithMeta
   at parse time. getnext/getprevious use direct index instead of
   linear scan over siblings. O(1) instead of O(siblings).

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

…gation

Major refactor to use simdxml's python-bindings-api branch directly:

- Drop IndexWithMeta entirely — no more custom parents, name_ids,
  child_positions, or unique_names. self_cell dependent is now just
  XmlIndex directly.
- .text uses upstream direct_text_first() — zero allocation
- .tail uses upstream tail_text() — O(log n) binary search instead
  of O(n) substring search through raw XML
- .getparent() uses upstream parent() — direct array lookup
- .getnext()/.getprevious() use upstream child_position() + child_at()
- __len__ uses upstream child_count() — zero allocation
- __getitem__ uses upstream child_at() — zero allocation
- child_tags/descendant_tags use upstream child_slice() — zero alloc
- attrib/keys/items use upstream attributes() — single-pass parsing
- Tag interning built from upstream name_ids/name_table (no rebuild)
- Cargo.toml: simdxml dependency points to git branch (temporary,
  will switch to crates.io release)

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

All numbers improved from upstream API integration:
- Parse: 2.2-3.1x vs lxml (was 1.4-1.8x)
- XPath text: 10-23x vs lxml (was 1.8-14x)
- XPath predicates: up to 42x vs lxml
- Traversal: 3-17x vs lxml via batch API
- No regressions vs lxml on any benchmark

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>