ESP-NN kernel does not allow for hybrid models (TFMIC-24)

[andresovela]

Running a model with a fully connected layer with different input / filter types fails due to the following assert in the ESP-NN kernel:

esp-tflite-micro/tensorflow/lite/micro/kernels/esp_nn/fully_connected.cc

Lines 69 to 70 in 7fd66e4

	TF_LITE_ENSURE_MSG(context, input->type == filter->type,
	"Hybrid models are not supported on TFLite Micro.");

In the normal kernel this condition is more relaxed:

esp-tflite-micro/tensorflow/lite/micro/kernels/fully_connected.cc

Lines 59 to 67 in 7fd66e4

	if ((input->type == kTfLiteFloat32 && filter->type != kTfLiteFloat32) ||
	(input->type == kTfLiteInt8 &&
	(filter->type != kTfLiteInt8 && filter->type != kTfLiteInt4)) ||
	(input->type == kTfLiteInt16 && filter->type != kTfLiteInt8)) {
	MicroPrintf("Input type: %s with filter type : %s not supported.",
	TfLiteTypeGetName(input->type),
	TfLiteTypeGetName(filter->type));
	return kTfLiteError;
	}

[vikramdattu]

@andresovela thanks for bringing it to our attention. I will update the port files by this weekend and release the code.

[vikramdattu]

@andresovela should be fixed with latest change: 30f24e6.
The latest commit is tagged with v1.3.1 and is pushed to component manager as well.