Add lazy "cudf" registration for p2p-related dispatch functions

[rjzamora]

While debugging some p2p failures in RAPIDS, I discovered that some workers occasionally hit the from_pyarrow_table_dispatch without having dask_cudf imported. This scenario is a bit hard to test, but the best way to avoid it is to have these p2p-related dispatch functions registered for a "lazy" dask_cudf import in dask.

[github-actions]

Unit Test Results

See test report for an extended history of previous test failures. This is useful for diagnosing flaky tests.

     15 files  ±0       15 suites  ±0   3h 19m 50s ⏱️ + 2m 11s
 13 107 tests ±0   12 173 ✅ ±0     930 💤 ±0  4 ❌ ±0 
162 263 runs  ±0  142 174 ✅ +4  20 085 💤  - 4  4 ❌ ±0 

For more details on these failures, see this check.

Results for commit 2b381ea. ± Comparison against base commit ba24c0a.

[quasiben]

Looks good. Is there a corresponding test we should add here or in the cuDF repo ?

[rjzamora]

Is there a corresponding test we should add here or in the cuDF repo

Test is a bit tricky and will be a modification of rapidsai/cudf#15469 - I'd like to avoid adding something fragile to gpuCI until the current suite is back in the green.

[quasiben]

Maybe we can come back to it after all PRs are merged in. Thanks @rjzamora