What does this PR do?

Fixes the issue where autocast_adapter_dtype=False was being ignored when using quantized models with BitsAndBytes.

Fixes #2889

Problem

When a model is quantized using BitsAndBytes (e.g., 4-bit quantization), LoRA adapters were always initialized with float32 dtype, even when:

• autocast_adapter_dtype=False was explicitly specified
• The model's compute dtype was set to float16

This caused unexpected behavior and potential performance/memory issues.

Solution

Added a _get_weight_dtype() helper method to the LoraLayer class that:

1. Checks for compute_dtype attribute (present in BitsAndBytes quantized layers)
2. Falls back to weight.dtype for regular layers
3. Uses this dtype when creating lora_A and lora_B Linear Layers in update_layer()

Testing

✅ Verified the fix works correctly with custom test script:

• Quantized model (4-bit with compute_dtype=float16) -> LoRA params are float16
• Non-quantized model (dtype=float16) -> LoRA params are float16
• Default behavior (autocast_adapter_dtype=True) still works as expected

✅ Ran 131 LoRA config tests locally - all passed