fix: offloading model layers to gpu and using flash_attn for gpu

[HeisenberG2575]

Context:

The _load_backbone functions loads the neutts-air backbone for inference. While this function works as intended for CPU inference, GPU inference is slowed down due to faults in code

Problem:

Considering that backbone_device is used to export the model to the appropriate device, "gpu" isn't a valid string for .to() hence users resort to the standard "cuda" or strings having "cuda" such as "cuda:0" as a prefix. The current if conditions

n_gpu_layers=-1 if backbone_device == "gpu" else 0
and
flash_attn=True if backbone_device == "gpu" else False
do not have the intended effect of offloading the model to GPU and using flash-attention since the string is checked for "gpu"

Fix:

• check string for "cuda" and "gpu" using python's inbuilt startswith method for strings ensures usage of "cuda" as backbone_device leads to intended behaviour
• checking for "gpu" alongside "cuda" ensures the change is not code-breaking for existing users

TODOs:

• include other device strings that may be utilized for GPU inference and support usage of flash_attn and use of n_gpu_layers with gguf