Support normalized operands in Add and Mul

[Johnkothapalli]

Summary

• route Add and Mul operands through GraphVariables.consume_operand_values(), matching Div and Sub
• support scalar, list, computed-column, and grouped-column operands with width-one broadcasting
• add executed-value regression coverage for the complete operand matrix and retain true width/type failures
• correct the MatMul module and coefficient-tensor wording

Closes #126.

Validation

• uv run pytest — 364 passed, 29 skipped
• uv run bash examples/test_examples.sh — all 11 examples passed (Git Bash used because .sh is not directly executable on Windows)
• uv run ruff check . — passed
• uv run ruff format --check . — 65 files already formatted
• uv run mypy — no issues in 47 source files
• uv run pre-commit run -a — Ruff hooks passed; the MyPy hook could not launch on native Windows because it hard-codes .venv/bin/mypy, so the same MyPy command was run directly above

AI assistance: Codex helped analyze the existing Div/Sub pattern, implement the change, and prepare tests. The full diff was independently reviewed and verified with the checks listed above.

[amol-]

@Johnkothapalli there seem to be conflict with the recently shipped improvements to the Optimizer. It should not be too hard to address those, it's primarily a matter of calling fold_operations in the right way