Handle dealloc in stream-ordered cudf-polars ops (#20467)

This updates cudf-polars' usage of CUDA streams to safely handle deallocation.

Consider the following sequence of operations:

1. Read some data on stream A
2. Read some data on stream B
3. Concat data from A and B on new stream C

cudf-polars currently ensures that C is downstream of `A` and `B` before doing the concat. But then our execution will typically drop all references to the data (on streams A or B), at which point Python's reference counting will call a `cudaFreeAsync` on streams A and B to free the memory used by the data from 1 and 2.

We need to ensure that this stream ordered free happens after the result from 3 (on stream C) is ready, and so we join stream C into each of A and B in each place where we previously just joined the streams.

Authors:
  - Tom Augspurger (https://github.com/TomAugspurger)

Approvers:
  - Richard (Rick) Zamora (https://github.com/rjzamora)
  - Lawrence Mitchell (https://github.com/wence-)

URL: #20467

Author: TomAugspurger