[FEA] Use optional runtime filtering optimizations in cudf-polars

[rjzamora]

Follow up to #22996 and #22997

As suggested by @wence-, we should encode pre-join filter optimizations as a runtime "hint", rather than a "required" logical operation. This way, the cost decision can be done at runtime, when the actors can observe real data properties.

Prerequisites:

• Add approx_distinct_count API to cudf_streaming #23522 (add approx_distinct_count streaming node for cardinality estimation)
• Remove bloom prefilter application in cudf-polars streaming #23523 (remove bloom pre-filtering)