Allow fp16 input tensor to apply with fp16 saved model

[deepkyu]

Problem

From @illian01

Compressor

• Current API gets input tensor as input_shape
  • Corresponding line
  • Generate input tensor with float32 with the given size
• Some models can be saved with half() (i.e. fp16), which cannot infer with float32 input tensor .
  • This model should give the example input tensor whose dtype is same with the model (float16)
• Need additional feature for users to select the data type of example input tensor
• Our back-end has already supported different input types of tensors, and it doesn't affect to compress the model

Related Link

Slack: https://nota-workspace.slack.com/archives/C040F65LSAJ/p1693546735327899

[Only-bottle]

@deepkyu @illian01님, 해당 이슈는 Compressor Backend에서 처리하였습니다.

해결한 방식은 아래와 같습니다.

• half의 경우 cpu에서는 사용할 수 없다는 것을 확인했고(관련 issue), float으로 변환해서 압축 후 half로 변환해서 제공하도록 처리하였습니다.