Fix predict_on_seqs OOM via batched Trainer.predict

[buyan-kh]

Summary

• Fixes Weird OOM in 1_inference.ipynb #154: predict_on_seqs eagerly moved the full one-hot batch onto the device and ran a single forward pass, which OOMs on large models (e.g. Borzoi in 1_inference.ipynb on 16GB VRAM).
• Aligns predict_on_seqs with predict_on_dataset / ISM_predict by using make_predict_loader + Lightning Trainer.predict so inference is batched.
• Adds optional devices, num_workers, batch_size, and precision kwargs (defaults: batch_size=1 for memory safety on large models).
• Keeps the device= keyword as a backward-compatible alias so existing tutorials (predict_on_seqs(..., device=0)) continue to work.

Test plan

• pytest tests/test_lightning.py::test_lightning_model_predict_on_seqs (CPU; covers single/multi seqs, devices=, device= alias, and batch_size)
• pytest tests/test_lightning.py::test_lightning_model_predict_on_dataset still passes
• Optional: re-run Borzoi predict_on_seqs from docs/tutorials/1_inference.ipynb on a 16GB GPU