Add low memory JSON reader for cudf.pandas

[galipremsagar]

Description

Fixes: #16122

This PR introduces low-memory JSON reading for cudf.pandas read_json.

Checklist

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

[mroeschke]

Just confirming that the concatenation technique here is generally the same as done in the parquet reader?

[galipremsagar]

yup

[mroeschke]

Just one non-blocking question

[mhaseeb123]

Just curious if we could use some logic such as _has_next() in PQ chunked reader to break this loop instead of this exception?

[vuule]

I don't like that we're catching the exception from a datasource here. The memory mapping is very much an implementation detail.
How expensive would it be to get the total source(s) size? Then we can loop until all of it is read.

[galipremsagar]

It is expensive, we basically have to seek to the end of file to get the size of all data sources. For remote data-sources it get's complicated to properly perform seek too.

[galipremsagar]

Just curious if we could use some logic such as _has_next() in PQ chunked reader to break this loop instead of this exception?

We basically call into libcudf layer for that, is there any such provision for json reader in libcudf?

[vuule]

FWIW, we already need the file size(s) to read JSON input(s). With the current implementation of the low memory JSON reader, we get the size of each input file for each byte range, so getting the sizes one more time to have a cleaner loop would not add much.

[galipremsagar]

@mroeschke @mhaseeb123 I just moved the chunked reader to pylibcudf to resolve the conflicts. Would you be able to take a look at the changes again?

[shrshi]

Thank you for working on this! I have just one comment based on the byte range reading behaviour we have seen in this issue.

[galipremsagar]

/merge