performance: improve performance for large template files

[JackByrne]

Description

This PR improves rendering performance for large docxtpl templates by removing a major XML manipulation bottleneck and reducing repeated per-render overhead in XML parsing, regex compilation, and Jinja2 environment setup.

The primary change avoids replacing the entire <w:body> element during rendering and instead mutates its children in place. On very large documents, this reduces render time from hours to minutes.

---

Problem

Rendering large templates (especially those with many tables and repeated data blocks) is extremely slow. Profiling showed that the dominant cost came from replacing the <w:body> element via root.replace(body, new_body) during render.

Replacing a large XML subtree triggers expensive libxml2 operations, including:

• Detaching the existing element
• Validating the incoming subtree
• Re-parenting and namespace reconciliation
• Fixing sibling links
• Re-homing or deep-copying nodes

For documents with millions of nodes, this results in effectively O(size of subtree) work and becomes the main performance bottleneck.

Additional overhead came from:

• Repeated compilation of the same regex patterns
• Recreating Jinja2 Environment objects on every render path
• Parsing XML into raw lxml elements that later need reconciliation with python-docx OXML elements

---

Current behavior

• Rendering replaces the <w:body> element entirely using root.replace(...)
• XML parsing in fix_tables() uses etree.fromstring(), producing raw lxml elements
• Regex patterns in patch_xml() and resolve_listing() are compiled on every render
• Jinja2 Environment objects are repeatedly constructed

On large real-world templates, this can result in render times of multiple hours.

---

Expected behavior

• Rendering should mutate the existing <w:body> element rather than replacing it
• XML should be parsed into python-docx OXML element types where possible
• Regex compilation and Jinja2 environment creation should be reused across renders
• Rendering output should remain unchanged

---

Fix

1. Mutate <w:body> children instead of replacing the element

• map_tree() now removes existing <w:body> children and appends rendered children in order
• This preserves the original body element identity used by python-docx
• Avoids expensive subtree replacement and reconciliation

2. Use parse_xml() in fix_tables() with safe fallback

• Primary parse uses docx.opc.oxml.parse_xml() to ensure OXML element subclasses
• Falls back to etree.fromstring(..., recover=True) to preserve robustness for malformed XML

3. Ensure new table grid columns use OXML elements

• New <w:gridCol> elements are created using OxmlElement and qn(...)
• Avoids mixing raw lxml elements with python-docx OXML elements

4. Precompile commonly used regex patterns

• Regexes in patch_xml() and resolve_listing() are compiled once and reused

5. Cache Jinja2 environments

• Reuses cached Environment instances (autoescape and non-autoescape)
• Avoids repeated environment creation across render paths

6. Header/footer fast path

• Skips rendering/parsing for headers and footers that contain no Jinja tags

---

Performance impact

On a large real-world template:

• Before: ~3 hours render time
• After: ~23 minutes render time

The majority of the improvement comes from avoiding <w:body> replacement.

---

Backward compatibility

• Output is intended to be equivalent to the previous implementation
• Behavior for malformed XML is preserved via fallback parsing
• Caller-provided Jinja2 environments are still respected

---

Notes

• All existing tests have been run and pass successfully.
• Manual verification was performed on large real-world templates.
• Rendering behavior and output are intended to remain unchanged; this PR focuses strictly on performance.