Benchmarking: wait for scalar bool output_valid on streaming models

[rameshkunasi]

Streaming models that gate feature extraction and classification subgraphs using Framer/Stacker and IF operators return kTfLiteOk with a 0D scalar boolean output_valid tensor set to false while internal circular buffers are warming up.

Update generic_model_benchmark to check for any 0D boolean output tensor and continue invoking until output_valid becomes true before logging steady-state profiler metrics and exiting the benchmark loop.

BUG=n/a

[ddavis-2015]

@rameshkunasi These changes are not sufficient to support all models reliably. I suggest adding makefile changes similar to how GENERIC_BENCHMARK_TENSOR_ARENA_SIZE is implemented. Lets call the new makefile var GENERIC_BENCHMARK_STREAMING_OUTPUT and the value it is set to is the MicroInterpreter output tensor index (0, 1, 2, ...). This allows specific builds of the benchmark tool for streaming models.

The IsOutputValid method should be adjusted to use the GENERIC_BENCHMARK_STREAMING_OUTPUT value, and have #if defined(GENERIC_BENCHMARK_STREAMING_OUTPUT) around that method and any use of the method.

[ddavis-2015]

@rameshkunasi Could you also make this change (not part of your original PR):

    if ((status != kTfLiteOk) && (static_cast<int>(status) != kTfLiteAbort)) {
      MicroPrintf("Model interpreter invocation failed: %d\n", status);
      return -1;
    }

Changed to:

    if (status != kTfLiteOk) {
      if (static_cast<int>(status) == kTfLiteAbort) {
        continue;
      }
      else {
        MicroPrintf("Model interpreter invocation failed: %d\n", status);
        return -1;
      }
    }

[ddavis-2015]

This PR is dependent on PR #3733