fix(data): order GPU representation reads after writers

[kevkrist]

Wait for recorded writer events before cloning, materializing, or converting GPU tables. Retain view ownership until asynchronous materialization completes, preventing cross-stream reads from racing producer work or freed buffers.

Regression coverage

This PR adds a deterministic cross-stream clone regression test in test_data_representation.cpp. It initializes the source allocation to zeros, blocks the producer stream before a queued write and the recorded writer event, then calls clone() on a separate consumer stream.

The test fails on main in two observable ways: clone() returns while the producer remains blocked, and the cloned buffer contains the stale zero pattern instead of the written value. The main implementation schedules the deep copy on the consumer stream without first waiting for the source writer event. This PR inserts the necessary stream dependency, so clone() remains blocked until the writer completes and copies the expected bytes.

The test commit also updates the remaining Catch2 v2 catch.hpp include to the Catch2 v3 catch_all.hpp header so the test target builds with the current dependency.

Validation

• pixi run build
• pixi run test (4/4 CTest targets)
• focused cuDF tests (537 assertions across 91 cases)
• repeated stream-ordering regression test
• negative control using the main clone() implementation: the regression test fails on premature completion and stale copied data
• changed-file lint and Doxygen
• Sirius integration tests

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.