Quantize all models to fp16

[WanliZhong]

This PR try to quantize all the models from float32 to float16. Accuracy evaluation results show they have almost same performance. Merged after OpenCV 4.8.0 released.

Models	Dataset	fp32 acc	fp16 acc
YuNet	WIDERFace	0.88437(E), 0.86560(M), 0.75030(H)	0.88425(E), 0.86558(M), 0.75018(H)
MobileNet V1	imagenet	0.6773(Top1), 0.8797(Top5)	0.6774(Top1), 0.88(Top5)
MobileNet V2	imagenet	0.6953(Top1), 0.8923(Top5)	0.6952(Top1), 0.8923(Top5)
PPResnet	imagenet	0.8238(Top1), 0.9615(Top5)	0.8238(Top1), 0.9615(Top5)
SFace	LFW	0.9940	0.9940
PPHumanSeg	mini supervisely	0.91642(IoU), 0.96558(Acc)	0.91641(IoU), 0.96557(Acc)
CRNN	ICDAR 2003	0.8166	0.8201
CRNN	IIIT5K	0.7433	0.7493

[WanliZhong]

@fengyuentau

1. Where the original float32 models place?
2. Should we name the new model xxx_fp16.onnx or just replace the fp32 models?
3. What's the difference among CRNN models (ch, cn, en; 2021, 2022, 2023)?

[fengyuentau]

Where the original float32 models place?

We need to discuss this. Probably end up in placing in a net drive.

Should we name the new model xxx_fp16.onnx or just replace the fp32 models?

Replace.

What's the difference among CRNN models (ch, cn, en; 2021, 2022, 2023)?

ch, cn, en stand for different charsets. Each charset contains different characters. You can find them in crnn.py.

[WanliZhong]

What's the difference between text_recognition_CRNN_EN_2021sep.onnx and text_recognition_CRNN_EN_2023feb.onnx? They have the same accuracy through evaluation in IIIT5K and ICDAR.