Improve the consistency of the PINNBench/Burgers benchmark

[Deathn0t]

The PINNBench/Burgers benchmark should be improved in the following way:

• [ ] adapt to deephyper==0.5.0 with receiving a job:RunningJob instead of config.
• [ ] @profile decorator should be used on the run-function.
• [ ] more metadata should be returned such as metrics, number of parameters and others see Combo model we should probably also define in the README all the metadata that Deep Neural Networks benchmark should return (see https://github.com/deephyper/benchmark/issues/4).
• [ ] TFKerasStopperCallback integration should be use to make it compatible with multi-fidelity see Combo hpo
• [ ] documentation should be added in the README of the benchmark: what is the problem, what are the scores, what are the run-times on some hardware, what is the baseline (neural network and related problem in general) etc...

[iamyixuan]

• [x] adapt to deephyper==0.5.0 with receiving a job:RunningJob instead of config.
• [x] @profile decorator should be used on the run-function.
• [x] more metadata should be returned such as metrics, number of parameters and others see Combo model we should probably also define in the README all the metadata that Deep Neural Networks benchmark should return (see https://github.com/deephyper/benchmark/issues/4).
• [ ] TFKerasStopperCallback` integration should be use to make it compatible with multi-fidelity see Combo hpo
• [x] documentation should be added in the README of the benchmark: what is the problem, what are the scores, what are the run-times on some hardware, what is the baseline (neural network and related problem in general) etc...

Finished the above tasks - TFKerasStopperCallback not implemented as the model in PINNBench is not complied with keras.

[Deathn0t]

Completed in https://github.com/deephyper/benchmark/pull/10