new feature : NAS-NLP-benchmark

[98may]

finish basic function, pass test with test_example.py.
Support query for both final and immediate training results.

[98may]

Usage:
0. cd nni/examples/nas/benchmarks

1. Install dependencies via pip3 install -r xxx.requirements.txt. xxx can be nlp.
2. Generate the database via ./xxx.sh. The directory that stores the benchmark file can be configured with NASBENCHMARK_DIR environment variable, which defaults to ~/.nni/nasbenchmark.

[98may]

TBD: check if query api should add an attribute called "reduction", which depends on if the paper's repo's "multi_run logs" is a superset of "single_run logs" .

[98may]

the paper's repo's "multi_run logs(m)" is not a superset of "single_run logs(s)", it has some dupilicates as the following data declares:
len(m) = 12343
len(set(m)) = 12343
len(s) = 14323
len(set(s)) = 14323
len(set(m) & set(s)) = 4114.

Now it is ready for review.

[ultmaster]

Please fix the pipelines.

[QuanluZhang]

@98may need to update doc accordingly

[QuanluZhang]

have you verified that the accuracy number can be reproduced with the authors' code?

[98may]

Not really. Do I need to double check that? If so, I will try to run the paper's repo's code.