Fix LoRA offload detection for alignment hooks (#10914)

[Bekhouche]

What does this PR do?

Fixes #10914.

LoRA loading previously treated every Accelerate AlignDevicesHook as evidence that sequential CPU offloading was enabled. Alignment-only hooks, such as those installed for quantized models, were therefore removed and followed by an unintended call to enable_sequential_cpu_offload(). This could move a GPU-resident model to CPU and significantly slow inference.

This change:

• only recognizes AlignDevicesHook instances with offload=True as sequential CPU offloading;
• inspects nested module hooks, preserving detection of genuine sequential offloading;
• preserves alignment-only hooks;
• adds regression coverage for both behaviors.

This updates the approach from the stale attempt in #8750 by checking the hook's explicit offload state instead of inferring it only from the hook type.

Tests

conda run -n conda3.12 env PYTHONPATH=src python -m pytest tests/lora/test_lora_loader_utils.py -q

8 passed