[converter] correct non-compliant axis handling in split and splitV ops

[LynnL4]

Hi,
This PR fixes the axis input for the split and splitV operations to match TensorFlow's definition, which requires a 0-D int32 tensor. The previous implementation incorrectly used an array, causing some model converters to fail.

FYI
https://tensorflow.google.cn/api_docs/python/tf/raw_ops/Split
https://tensorflow.google.cn/api_docs/python/tf/raw_ops/SplitV

[CLAassistant]

All committers have signed the CLA.

[peterjc123]

Thanks for your PR. Let's wait for the CI to finish

[peterjc123]

According to the docs here, it seems that tfl.split_v needs a 0-d tensor as dim, but it doesn't to be mentioned for tfl.split.

[LynnL4]

According to the docs here, it seems that tfl.split_v needs a 0-d tensor as dim, but it doesn't to be mentioned for tfl.split.

From the case of the code, it looks like the implementation of the two operators regarding this part is the same. axis takes only one value.

https://github.com/tensorflow/tflite-micro/blob/d3475aa77621545b7b0ef585f38fb8634c7ee766/tensorflow/lite/micro/kernels/split.cc#L89

https://github.com/tensorflow/tflite-micro/blob/d3475aa77621545b7b0ef585f38fb8634c7ee766/tensorflow/lite/micro/kernels/split_v.cc#L92

https://github.com/tensorflow/tensorflow/blob/631973738040e1fdceb02584c5a07317f333d732/tensorflow/lite/kernels/split.cc#L54
https://github.com/tensorflow/tensorflow/blob/631973738040e1fdceb02584c5a07317f333d732/tensorflow/lite/kernels/split_v.cc#L176

[peterjc123]

@LynnL4 Yes, it looks like TFL core doesn't care about the shape of the dim tensor. We could adapt to 0-d tensors if other projects do this as well. Please fix the unit tests so that I can take the changes in.

[LynnL4]

@LynnL4 Yes, it looks like TFL core doesn't care about the shape of the dim tensor. We could adapt to 0-d tensors if other projects do this as well. Please fix the unit tests so that I can take the changes in.

All I know is that the ST EDGE AI conversion tool has this limitation. Also on the arm vela conversion tool it converts fine whether it's 0-D or not.

[LynnL4]

@peterjc123 Bugs have been fixed according to CI