c.parallel: investigate making ptx-json generate valid PTX

[griwes]

Currently, ptx-json emits invalid PTX, which means that regardless of whether we need to do a two C++ frontend pass compilation or not when building kernels, we have to do it. #5355 serves as a proof of concept for extracting data from PTX generated from LTOIR, and adapting that technique in c.parallel would allow us to eliminate one of the C++ frontend passes, speeding the build step considerably. Investigate making this possible.

[griwes]

PTX doesn't actually have strings, so that exact method would require more postprocessing than desired. However, it looks like treating the final cubin as the blob to search for the json in is promising.

[griwes]

Proof of concept for the new approach: https://www.godbolt.org/z/W9Wzsv5xK. Looks like doing a single pass compilation for reduce cuts the runtime of tests from 90s to 60s.