BUG: Pip 20.3 is causing Linux py38 np dev pipeline to fail

[phofl]

The first pipeline to fail was 2a115d5 which also was the first pipeline using pip 20.3

This seems to be a problem with scipy version.

@jreback @jbrockmendel @jorisvandenbossche

Last pipeline to work:
https://dev.azure.com/pandas-dev/pandas/_build/results?buildId=49130&view=logs&j=eab14f69-13b6-5db7-daeb-7b778629410b&t=71ac627c-7b38-59b1-014b-81753001ec5e

First fo fail:
https://dev.azure.com/pandas-dev/pandas/_build/results?buildId=49159&view=logs&j=eab14f69-13b6-5db7-daeb-7b778629410b&t=71ac627c-7b38-59b1-014b-81753001ec5e&l=333

[jorisvandenbossche]

See pypa/pip#9186

[jorisvandenbossche]

So once this is fixed upstream, we should revert the pin.

[pradyunsg]

Noting that, from my PoV, the underlying issue here is that pip's getting wheels with versions that are not valid (see pypa/pip#9186 (comment)) and this is the first version of pip where it rejects such wheels by default. The fix probably needs to happen wherever the wheels are being generated.