Replace deprecated torch.cuda.amp API with torch.amp equivalents

[xyf5432]

Fixes #2420

Summary

torch.cuda.amp.GradScaler has been deprecated since torch 2.3 and torch.cuda.amp.autocast since torch 2.4, both scheduled for removal. This PR migrates the eight __main__ self-test usages to torch.amp, which provides the same APIs (GradScaler since torch 2.3, autocast since torch 2.0) and is compatible with the torch ≥ 2.6.0 floor the README already declares. The IPEX hijack aliasing in library/ipex/hijacks.py already targets torch.amp, so it only gets a clarifying comment.

Changes:

• library/sdxl_original_unet.py (2) — torch.cuda.amp.GradScaler(enabled=True) → torch.amp.GradScaler("cuda", enabled=True); with torch.cuda.amp.autocast(enabled=True): → with torch.amp.autocast("cuda", enabled=True):
• library/sdxl_original_control_net.py (2) — same migration, keeping dtype=torch.bfloat16
• networks/control_net_lllite.py (2) — same migration
• networks/control_net_lllite_for_train.py (2) — same migration, keeping dtype=torch.bfloat16
• library/ipex/hijacks.py (0) — comment only: the torch.cuda.amp aliasing in ipex_hijacks() (IPEX-only path) replaces torch.cuda.amp with the torch.amp / torch.xpu.amp implementations, so deprecated names are never invoked; the comment documents why the references stay

All eight changes are inside if __name__ == "__main__": self-test blocks (developer smoke tests), none on the main training path (train_network.py / sdxl_train.py / train_util.py have zero torch.cuda.amp references).

Type of change

• Bug fix (non-breaking change which fixes an issue)

Validation

Verified on torch 2.11.0 (CPU build) with warnings.simplefilter("error", FutureWarning):

• torch.amp.GradScaler("cuda", enabled=True) constructs and runs warning-free.
• torch.amp.autocast("cuda", enabled=True) and the dtype=torch.bfloat16 variant enter cleanly (autocast is disabled when CUDA is unavailable, by design).
• Control group: torch.cuda.amp.GradScaler(enabled=True) raises FutureWarning under the same filter — the migration is effective and the filter is sensitive.
• IPEX hijack pattern verified: reading torch.cuda.amp.custom_fwd (attribute access, not a call) emits no warning, and after torch.cuda.amp = torch.amp the alias resolves to the torch.amp implementation.
• py_compile passes on all five touched files.

User impact

No behavior change; the FutureWarning emitted when running the self-test blocks on the documented torch ≥ 2.6.0 install path is eliminated, as is the risk of breakage when torch removes torch.cuda.amp.

Notes for reviewers

• torch.amp.autocast exists since torch 2.0 and torch.amp.GradScaler since torch 2.3, so no version branching is needed on the declared floor (README: "PyTorch 2.6.0 or later is required").
• Same migration as ultralytics/yolov5#13244 and huggingface/lerobot#3167.