Use cached input to speed up tensor initialization in tflite benchmark model.

This CL changes how data inputs are generated for benchmark_tflite_model:
Previous: We random initialize all the input during each run, which is time consuming. And also contribute to the mysterious 'other' tag in the benchmark result.
After: We will initialize input only once before inference, and in each run reset the input from the cache.

After this change each run will have exactly the same data input, which shouldn't be a big concern (especially for larger models).

PiperOrigin-RevId: 236875318