[BUG] Invalid characters in URI query are handled differently when reading from GPU

[hyperbolic2346]

Describe the bug
Invalid characters such as extended ascii encoded in hex format, eg %9F, are handled differently when parse_uri is returning a query. The GPU kernel will leave these values as-is and return an extended ascii character, but this is invalid UTF-8 and the CPU replaces these characters with � (0xefbfbd).

Steps/Code to reproduce bug
Call parse_uri with a string such as "http://www.nvidia.com/object.php?object=ะก-Ð%9Fะฑ".

Expected behavior
It would be ideal to produce bit-for-bit compatible results as the CPU. It is unknown what other characters are replaced.

[hyperbolic2346]

This seems very similar to #9560 and the solutions may be the same depending on how CSV parsing is handled.

[hyperbolic2346]

This may be a Java difference from Spark. Spark shell is not unescaping these values.

[hyperbolic2346]

Spark calls the raw API for the Java URI functions which doesn't decode. This means we can simply pass the hex-encoded data as-is and there is no translation necessary.