fix: use float32 for Qwen image rotary embeddings

[huntcsg]

Problem

When using torch.compile with bfloat16 inference, Qwen image models fail with:

No backend can handle 'apply_rope1': eager: freqs_cis: dtype torch.bfloat16 not in {torch.float32}

Root Cause

1. Commit 4cd88186 ('Use single apply_rope function across models') refactored all models to use a shared apply_rope1 function and correctly set Qwen's image_rotary_emb to .to(torch.float32)

2. Commit c4a6b389 ('Lower ltxv mem usage') inadvertently reverted Qwen back to .to(x.dtype) (bfloat16), breaking compatibility with the shared apply_rope1 function

3. The apply_rope1 function uses addcmul_ in-place operations that fail under torch.compile when freqs_cis is bfloat16

Fix

This restores the correct behavior by ensuring image_rotary_emb is always float32 for the Qwen model, matching the expected dtype for apply_rope1.

Testing

Tested with Qwen Image Edit workflows on cloud deployment.

[comfyanonymous]

This should be fixed if you update your comfyui with the requirements.txt