fix: skip incompatible LoRA weights

[leejet]

Summary

• skip shape-incompatible LoRA weights with a warning instead of aborting
• validate runtime linear LoRA down/mid/up dimensions against the base weight
• track skipped tensors separately and avoid repeated warnings or misleading applied counts

Related Issue / Discussion

Fix #1796.

Additional Information

N/A

Checklist

• I have read and confirmed this PR follows the contribution guidelines.

[masamaru-san]

FYI: I use only the SD 3.5 series but when using sd-cli with --lora-apply-mode at_runtime, this PR causes many tensors in LoRA to be skipped. For example, the 8‑step LoRA for the Medium‑Turbo model.

... should I create new issue?

[stduhpf]

I beleive it's because of the split qkv loras, the .add_q_proj.weight.lora_A gets renamed as .qkv.weight.lora_down, and is supposed to be concatenated with .qkv.weight.1.lora_down (add_k_proj) and .qkv.weight.2.lora_down (add_v_proj), so it doesn't match the size of the full .qkv.weight matrix alone.