feat: add Qwen3-VL tensor parallel

[kcz358]

Summary

• add tp_degree mesh setup and Qwen3-VL text-decoder tensor parallelization
• shard Qwen3-VL attention q/k/v/o and MLP gate/up/down projections over the TP mesh
• account for TP in FSDP2 throughput/MFU token de-duplication
• document Qwen3-VL TP usage in README, model docs, and examples

Verification

• User ran training smoke with TP enabled
• pre-commit hooks on commits: black/isort passed