Place metric functions for BLEU and Rogue on correct devices when using multiple GPUs

[arnavgarg1]

The issue was that the metric function wasn't being moved/placed on the right device, leading to a weird behavior for these metrics using the response prediction key because the inputs that are passed in are not tensors, they're lists of strings. However, it seems like these metric functions need to moved to CUDA (instead of staying on the CPU) so that when the metric_fn.compute() call is called to gather evaluation metric summaries, it does not run into this error:

RuntimeError: Tensors must be CUDA and dense

Tested successfully with:

1. Only CPU machine
2. Multi-GPU Quantized training (4 GPUs)
3. Multi-GPU DeepSpeed Stage 3 training (4 GPUs)

[github-actions]

Unit Test Results

       6 files  ±0         6 suites  ±0   53m 16s ⏱️ + 2m 5s
2 807 tests ±0  2 793 ✔️ ±0  12 💤 ±0  2 ❌ ±0 
2 847 runs  ±0  2 824 ✔️ ±0  21 💤 ±0  2 ❌ ±0 

For more details on these failures, see this check.

Results for commit e7d0f6f. ± Comparison against base commit 4af5331.