Expose stream-ordering in scalar and avro APIs

[shrshi]

Description

Contributes to #13744

Replaces conversion operators in derived classes of cudf::scalar with stream-ordered get_value(stream) member function.
Adds stream parameter to cudf::io::read_avro

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[bdice]

RMM just calls this value. We often try to avoid “get” in our function names.

Suggested change

	T get_value(rmm::cuda_stream_view stream = cudf::get_default_stream()) const;
	T value(rmm::cuda_stream_view stream = cudf::get_default_stream()) const;

[bdice]

Wait, doesn’t value already exist just below here? We may not need get_value at all. Maybe we just need to delete the conversion operator since it does not take a stream.

[shrshi]

Good point - I chose to introduce another function here because value can return either host or device data depending on the type. string_scalar::value returns a cudf::string_view on the device, while fixed_point::value and fixed_width::value copy to host. I was also unsure of using fixed_point::value directly since the conversion operator returns the fixed_point_value instead of the unscaled value.

[shrshi]

I'm hesitant to modify string_scalar::value to return a host string view since it is used in several places that directly access the scalar on device. One option is to rename this function to string_scalar::d_value and introduce another string_scalar::value that returns a std::string?

[davidwendt]

The string_scalar::value() should return a cudf::string_view which is possible from host or device since it just wraps a pointer and a size held by the string_scalar. No device code is needed. Actually the stream parameter is not used at all.
To return a host std::string one would use the string_scalar::to_string() function.

[davidwendt]

I agree with Bradley we should not have the get_value() functions and just keep the value() ones.

[shrshi]

Thanks David, Bradley. I've retained the value functions and got rid of the conversion operators.

[bdice]

Please refactor the tests like this:

Suggested change

	auto lhs_h = static_cast<ScalarType const&>(lhs).get_value(cudf::get_default_stream());
	auto lhs_h = static_cast<ScalarType const&>(lhs).value(cudf::get_default_stream());

[shrshi]

I've added a scalar_host_value function that checks which scalar class is passed since the conversion operator in string_scalar and fixed_point_scalar called to_string and fixed_point_value respectively.

[bdice]

One style request, otherwise LGTM!

[galipremsagar]

@shrshi I cancelled the most recent workflow to free up resources to unblock all of cudf CI for this PR: #17771

I'll rerun once #17771 is merged.

[vuule]

looks good, just one typo to fix

[vyasr]

Marking as DO NOT MERGE until we've had some discussion with users.

[nvdbaranec]

I'm having issues building spark-rapids-jni with this branch - cmake related stuff. Would it be possible to get an up to date merge of 25.02 pulled into here?

[nvdbaranec]

All spark tests pass. Good to go.

[shrshi]

/merge