Optimize Tiny YoloV2 model by fusing the conv + batchNormalization

[Honry]

• Fuse conv + batchNormalization would bring some performance benefit
• Convert weights to fp16 for GPU and NPU backends

[Honry]

@huningxin, PTAL, though Tiny Yolo doesn't work on NPU now, this optimization can still bring some benefit for GPU.

I am thinking of adding data type option to support both fp32 and fp16 and disabling the unsupported models for each backend in a follow up.

If this one looks good, please help merge the test-data change at webmachinelearning/test-data#21 firstly.

[huningxin]

though Tiny Yolo doesn't work on NPU now

Does it make sense to hide TinyYolo when user select NPU until it is supported?

[Honry]

Does it make sense to hide TinyYolo when user select NPU until it is supported?

I have a follow-up PR to support both fp16 and fp32, will hide TinyYolo fp16 for NPU.

[huningxin]

LGTM