Delegate kernel selection for int8 conv to our library

- kelvin::opt::conv_per_channel_b8 now selects from several optimized
  implementations, or falls back to the generic scalar implementation.

Change-Id: I26ea73368ea43728575ff1557330326399ebacf7

tensorflow/lite/micro/kernels/kelvin/conv.cc

diff --git a/tensorflow/lite/micro/kernels/kelvin/conv.cc b/tensorflow/lite/micro/kernels/kelvin/conv.cc
index 28ec1e8..2e15dd7 100644
--- a/tensorflow/lite/micro/kernels/kelvin/conv.cc
+++ b/tensorflow/lite/micro/kernels/kelvin/conv.cc
@@ -136,13 +136,7 @@
         }
         case kTfLiteInt8: {
           const auto params_q = ConvParamsQuantized(params, data);
-          bool opt = !(params_q.padding_values.width > 0 ||
-              params_q.padding_values.height > 0 ||
-              params_q.dilation_width_factor > 1 ||
-              params_q.dilation_height_factor > 1);
-          auto fn = kelvin::opt::conv_per_channel_b8;
-          if (!opt) fn = reference_integer_ops::ConvPerChannel;
-          fn(
+          kelvin::opt::conv_per_channel_b8(
               params_q, data.per_channel_output_multiplier,
               data.per_channel_output_shift,
               tflite::micro::GetTensorShape(input),