Parse XML tool-use repl blocks in find_code_blocks

[MohammadHijjawi97]

Fixes #179

Claude models often emit their native XML tool-use syntax instead of repl fences:

<function_calls>
<invoke name="repl">
<parameter name="code">
print(len(context))
</parameter>
</invoke>
</function_calls>

find_code_blocks() only matched fences, so none of that code was executed and the model went on to hallucinate the outputs.

Changes:

• find_code_blocks() now falls back to parsing <invoke name="repl"><parameter name="code">...</parameter></invoke> blocks when no repl fence is present, which is the fix proposed in the issue. Fenced parsing and ordering are unchanged, and fences take precedence when both appear.
• Fixed the docstring: the function returns an empty list, not None.
• Added unit tests in tests/test_parsing.py for single and multiple XML blocks, ignoring non-repl invokes, and fence precedence.

Tests: pytest tests/test_parsing.py tests/test_imports.py (55 passed, 3 skipped); ruff check / ruff format --check clean. The XML tests fail without the change.

Note: #175 adds a None guard in the same function, so whichever lands second may need a trivial rebase.