fix(dreambooth): batch size mismatch with --with_prior_preservation in flux2 scripts

[agarwalprakhar2511]

Summary

When --with_prior_preservation is enabled in the Flux2 dreambooth LoRA training scripts, the prompt embedding repeat logic double-counts the batch size, producing a shape mismatch against the latent tensor.

Root cause: collate_fn appends class prompts to the instance prompts list (doubling len(prompts)), but prompt_embeds is already doubled earlier via torch.cat([instance_embeds, class_embeds]). Using the full len(prompts) as the repeat count produces 4 embeddings for 2 latents at batch_size=1.

Fix: Use len(prompts) // 2 when args.with_prior_preservation is active, so the repeat count matches the number of unique prompt groups rather than the doubled collated list.

Applied to all three affected scripts:

• examples/dreambooth/train_dreambooth_lora_flux2_klein.py
• examples/dreambooth/train_dreambooth_lora_flux2.py
• examples/dreambooth/train_dreambooth_lora_flux2_klein_img2img.py

Test plan

Dimension trace verified for all four scenarios:

batch_size	with_prior_preservation	len(prompts)	prompt_embeds pre-repeat	Repeat count (old → new)	Result (old → new)
1	True	2	[2, seq, H]	2 → 1	[4, seq, H] MISMATCH → [2, seq, H] OK
2	True	4	[2, seq, H]	4 → 2	[8, seq, H] MISMATCH → [4, seq, H] OK
1	False	1	[1, seq, H]	1 → 1	[1, seq, H] OK → [1, seq, H] OK
2	False	2	[1, seq, H]	2 → 2	[2, seq, H] OK → [2, seq, H] OK

Fixes #13292