refactor: unify cpu gpu into device

[bwanglzu]

Resolve inconsistency between fit end get_model, in fit we use cpu, in get_model we use gpu. Better to align with Docarray device='cpu' or device='cuda'.

This will not be a breaking change for internal users.

This will be a breaking change for ppl install finetuner-full when calling get_model.

A follow up ticket has been created to remove cpu in 0.7.0 #578

# training time
run = finetuner.fit(
    ...,
-   cpu=False,  # cpu is still supported, no breaking change, we sunset `cpu` after 0.7
+   device='cuda', # note, now `cuda` is the default device.
)

# inference time
inference_engine = finetuner.get_model(
     ...,
-    gpu=True, 
+    device='cuda',  # note, `cpu` is the default device unless user specify `cuda`.
)

---

• This PR references an open issue
• I have added a line about this change to CHANGELOG

[github-actions]

📝 Docs are deployed on https://ft-refactor-device--jina-docs.netlify.app 🎉