Benchmark gpt-tfjs integration

[JulienVig]

Benchmark and report some information on what to expect and the row for improvement:

• [ ] number of training iterations needed on an example dataset, expected loss/perplexity, example of text generation after training
• [ ] number of tokens/sec
• [ ] model sizes
• [ ] memory requirements
• [ ] compare numbers with ONNX

#656 should be addressed first