Run cudf-polars wheels unit tests with more than 1 process

[mroeschke]

Description

Follow up to #19980

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[bdice]

Looks fine. Please report the before/after times.

[vyasr]

Can we add a note about why session scope here is important here?

[mroeschke]

Looks fine. Please report the before/after times.

Here's the raw pytest run times comparing the CUDA 12.9 cudf_polars wheels test CI job f4554de from this branch to the same job in #20141 as <minute>.<seconds> (the similar CUDA 13.01 job shows a similar magnitude decrease)

Before (#20141)

Polars version	in-memory	streaming	streaming + small block	streaming + distributed
1.28.1	1.33	1.33	1.47	2.04
1.29.0	1.34	1.33	1.47	2.10
1.30.0	1.35	1.36	1.49	2.15
1.31.0	1.39	1.37	1.51	2.10
1.32.3	1.45	1.48	2.05	2.23

After (f4554de)

Polars version	in-memory	streaming	streaming + small block	streaming + distributed
1.28.1	1.04	1.04	1.13	0.32
1.29.0	1.05	1.05	1.13	0.33
1.30.0	1.05	1.05	1.13	0.34
1.31.0	1.05	1.05	1.14	0.33
1.32.3	1.08	1.09	1.17	0.37

The overall job time has decreased from ~43 minutes to ~22 minutes.

[vyasr]

That's great! Glad to see parallelism getting us enough improvement here to keep this as is.

[mroeschke]

/merge