Align matrix and expand emission with PTO model

[zhoubot]

Summary

• emit explicit zero padding for Matrix B.DATR so PTO CUBE legality accepts generated bundles
• describe TROWEXPAND/TCOLEXPAND with destination geometry while preserving relative source bindings
• add contract and compile fixture coverage for the corrected forms

Evidence

• python3 test/test_v058_engine_contract.py (36 passed)
• python3 tools/generate_engine_docs.py --check
• real SuperNPUBench FP32 TROWEXPAND/TCOLEXPAND ELFs pass the ASL model
• real TMATMUL ELF disassembles as B.IOT t#1, m#1, ... ->u and B.DATR FP32, byte0, Zero

Dependency

The four-operand B.DATR spelling requires the companion LLVM change in LinxISA/llvm-project#78. This PR is intentionally not merged yet.