ARROW-2917: [Python] Use detach() to avoid PyTorch gradient errors

[alok]

detach() doesn't copy data unless it has to and will give a RuntimeError if the detached data needs to have its gradient calculated.

[kou]

Thanks for your contribution.
Can you open a JIRA ticket and prepend ARROW-XXX: to the title?
https://issues.apache.org/jira/browse/ARROW

https://github.com/apache/arrow/blob/master/.github/CONTRIBUTING.md#how-to-contribute-patches

To contribute a patch:

1. Break your work into small, single-purpose patches if possible. It’s much harder to merge in a large change with a lot of disjoint features.
2. Create a JIRA for your patch on the Arrow Project JIRA.
3. Submit the patch as a GitHub pull request against the master branch. For a tutorial, see the GitHub guides on forking a repo and sending a pull request. Prefix your pull request name with the JIRA name (ex: ARROW-423: Define BUILD_BYPRODUCTS for CMake 3.2+ #240).
4. Make sure that your code passes the unit tests. You can find instructions how to run the unit tests for each Arrow component in its respective README file.
5. Add new unit tests for your code.

[codecov-io]

Codecov Report

Merging #2311 into master will increase coverage by 2.38%.
The diff coverage is 0%.

@@            Coverage Diff             @@
##           master    #2311      +/-   ##
==========================================
+ Coverage   84.38%   86.76%   +2.38%     
==========================================
  Files         293      237      -56     
  Lines       44786    42045    -2741     
==========================================
- Hits        37791    36482    -1309     
+ Misses       6964     5563    -1401     
+ Partials       31        0      -31

Impacted Files	Coverage Δ
python/pyarrow/serialization.py	81.39% <0%> (ø)	⬆️
go/arrow/array/numeric.gen.go
go/arrow/array/numericbuilder.gen.go
go/arrow/type_traits_boolean.go
go/arrow/memory/memory_sse4_amd64.go
go/arrow/math/int64.go
rust/src/memory_pool.rs
go/arrow/datatype_binary.go
go/arrow/array/booleanbuilder.go
go/arrow/math/float64_sse4_amd64.go
... and 47 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update eaa6053...f8e298f. Read the comment docs.

[alok]

@kou I'd rather not make a JIRA account for a one word change. Would you mind adding the patch? It does pass the tests.

[wesm]

Is there any way to trigger the problem that this fixes? Having an ASF JIRA account is a good idea; at some point you may find yourself contributing a larger patch to an Apache project.

[alok]

@wesm Here's a script that triggers it for me: https://gist.github.com/c683d79cbc4b5c51282df983d81856ae

[alok]

@wesm @kou Created JIRA issue and updated title.

[wesm]

I added a unit test; all you had to do was try serializing a tensor with requires_grad=True.

[wesm]

+1

[alok]

@wesm Thanks for adding the test that I really should've.