Remove Path key/subscript to save GPU memory usage to improve perf

[res-life]

Remove Path key/subscript to save GPU memory usage to improve perf.

I found less GPU stack memory usage will help on perf.
I decreased the max path depth from 16 to 8, and get perf improvement.

constexpr int max_path_depth = 8;

And I optimized the main flow logic to reduced query paths:

• key and named to one path => named
• subscript and index/wildcard to one path => index/wildcard
  From the definition of JSON query path, key and named always ocurr in the same time, and also subscript and index/wildcard are always occur in the same time. So it's safe to reduce (key, named) to (named), and it's safe to reduce (subscript, index/wildcard) to (index/wildcard)

Note: If query path is a single wildcard , e.g.: $.* , Spark returns null, so the path is equivalent to invalid path.

performance test

Customer	Base line(1st, 2nd, 3rd) cudf legacy	After this PR	speed
customer 1	231500 ms, 232760 ms, 232647 ms	275719 ms, 272898 ms, 274407 ms	about 84%

[res-life]

Some IT cases are not passed:

./integration_tests/run_pyspark_from_build.sh -s -k test_get_json_object
========= 9 failed, 58 passed, 27086 deselected, 9 warnings in 18.72s ==========

[res-life]

All IT cases passed after applying two PRs:

• Add short circuit path for get-json-object when there is separate wildcard path spark-rapids#10734
• This PR

[revans2]

Can we please split up these two changes so that we can evaluate the performance of each of them separately? I would really like to see max_path_depth be turned into a template parameter for the kernel. That way we can have a few different kernels and select the right one depending on the path done. That or we can do like @nvdbaranec did on the original version and set the amount of shared memory that would be needed as part of launching the kernel. But that can be follow on work.

[res-life]

Can we please split up these two changes so that we can evaluate the performance of each of them separately?

Now, we combine [Key, Named] 2 paths (e.g.: .name1) to a single [Named] path, so the change from 16 to 8 is not a tuning.
And:
[Subscribe, Index] e.g.: [1] => [Index]
[Subscribe, Wildcard] e.g.: [*] => [Wildcard]
No need to separate to two PRs.

That or we can do like @nvdbaranec did on the original version and set the amount of shared memory that would be needed as part of launching the kernel. But that can be follow on work.

OK, will add in a follow-up.

[res-life]

build