[Dataset] Bump numpy >=1.20 dependency

[wuisawesome]

Why are these changes needed?

We need to ensure our numpy version is greater than 1.20. In 1.19 and under, the following code doesn't work

import pandas as pd
import torch
pd.DataFrame([torch.arange(27).reshape(3,3,3) for _ in range(10)])

Right now, this is pretty fundamental to how we build datasets of tensors, which is common when doing last-mile data preprocessing (i.e. a preprocessed image is most likely represented as a tensor).

Note that 1.20 dropped support for python 3.6 so the best we can do is enforced this versioning when using newer python versions.

Related issue number

Closes #20258

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[jjyao]

I'm wondering what's the difference between data extras and https://github.com/ray-project/ray/blob/master/python/requirements/data_processing/requirements_dataset.txt

[wuisawesome]

Data extras is what's actually installed when you do pip install ray[data]. Those requirements files should be for additional ray developer add-ons.

[ericl]

@wuisawesome can you elaborate on what "doesn't work" means? Can we raise a better error message in this condition?

[wuisawesome]

Yeah by "doesn't work" I mean it throws this error

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/ubuntu/.local/lib/python3.8/site-packages/pandas/core/frame.py", line 711, in __init__
    mgr = ndarray_to_mgr(
  File "/home/ubuntu/.local/lib/python3.8/site-packages/pandas/core/internals/construction.py", line 304, in ndarray_to_mgr
    values = _prep_ndarray(values, copy=copy)
  File "/home/ubuntu/.local/lib/python3.8/site-packages/pandas/core/internals/construction.py", line 540, in _prep_ndarray
    values = np.array([convert(v) for v in values])
ValueError: only one element tensors can be converted to Python scalars

Note that other tensors (like ndarray) also don't work, but throw different error messages. To make better error messages we could try to swallow all errors and tell people to upgrade their numpy version (seems a little scary) or start inspecting the contents of the dataset and doing special case checks for the various tensor types if numpy < 1.20 on python 3.6.

[wuisawesome]

test_output is broken on master