Make tables spillable by default

[abellina]

Closes #7672

This depends on rapidsai/cudf#13260 (rapidsai/cudf#13180 overarching issue)

The idea behind this is to be able to add a Table as an object into the spillable store. The Table will not be made contiguous until that is needed at spill time (to host memory), which is done using chunked_pack. Once it has gone through chunked_pack the table will be reconstituted as a regular RapidsDeviceMemoryBuffer with a single contiguous allocation.

It is draft because I still need to clean it up some and I need to add unit tests specific to this scenario. I also did a bunch of cleanup today and I have not run the latest, so I could have bugs here. But nevertheless, I think it is worth starting to take a look at it.

It won't pass the build because the cuDF PRs are not in yet, and I have one test failure that I know of.

[revans2]

Mostly nits and comment requests. But I want to spend some more time looking at the tests.

[jbrennan333]

Overall this is excellent work @abellina! I will likely do another pass, but so far my comments all overlap with those from @revans2.

[abellina]

Updating the full NDS benchmark with rapidsai/cudf#13260 + rapidsai/cudf#13278 + #8264 (this), I am seeing a ~4% improvement vs the last nightly.

I looked at one of the queries (q95) which becomes ~14% faster. Before the change ~7% of kernel time was spent in this function, whereas now it's less than 0.1%.

Name = benchmark
Means = 417757.4, 401496.8
Time diff = 16260.600000000035
Speedup = 1.0404999491901306
T-Test (test statistic, p value, df) = 9.948692331384382, 8.820392981431996e-06, 8.0
T-Test Confidence Interval = 12491.56085504279, 20029.63914495728
ALERT: significant change has been detected (p-value < 0.05)
ALERT: improvement in performance has been observed

[abellina]

working on review comments locally and will push but this won't be merge-able until Monday.

[abellina]

build

[abellina]

build

[abellina]

build

[abellina]

Ok, @revans2 the reason for c3c2179 was a change I had made in the test locally and forgotten to undo it. It should be a single close because the code throws a retry and split within a retry (this tests the specific exception), and so every batch that was open was told to close. I do not remember why I had changed this locally, but overall it removes the whole file from the unit test changes.