fix(lora): load ai-toolkit DoRA magnitudes across architectures

[zjn20030811]

Summary

ai-toolkit DoRA adapters store the output-dimension magnitude as either .magnitude or .lora_magnitude_vector.weight. The LoRA converters previously split these keys with a fallback parser, producing invalid layer groups or an unsupported format error. This change recognizes both suffixes across the Wan, Z-Image, Anima, Qwen Image, and FLUX conversion paths and maps them to the existing dora_magnitude representation.

The FLUX fused-QKV path now carries the magnitude (and existing input-dimension scale) into each output slice, preserving the correct orientation during conversion. Anima LoKR combinations strip either DoRA magnitude convention consistently.

Related Issues / Discussions

Closes #9558

QA Instructions

• Added non-square output-dimension regression coverage for both suffix spellings in each affected architecture.
• Added fused-QKV FLUX coverage to verify output-channel magnitude slicing.
• Ran the focused conversion suite: 133 passed.
• Ran ruff check, ruff format --check, and git diff --check on all changed files.

Merge Plan

Normal backend merge; no schema or database changes.