Performance optimizations for QSS chemistry

[baperry2]

Draft PR with optimizations from the NREL/NOAA/NCAR Open Hackathon:

1. Remove separate function that computs/stores k_f - just compute it on an as-needed basis (1% performance gain)
2. Reuse values of qf and qr computed during QSS coefficient calculation instead of recomputing (10% cumulative performance gain)

Nsight Compute profiles of the reaction kernel, before, after (1), and after (2), respectively:

Another optimization that was tried by @SreejithNREL was adding #pragma unroll to various species loops - that didn't have an effect as the compiler seems to being doing it automatically without the pragma.

TODOs:

• Verify correctness of commit that re-used qf and qr. That commit lead to diffs greater than machine precision, but small enough for still accurate simulations
• Modify these changes so they don't interfere with symbolic analytical Jacobian stuff
• Revert commit that removes scope because that didn't have an effect
• Regenerate all chemical mechanisms (currently only dodecane_lu_qss is updated, and without analytical jacobian support)

Thanks @dappelha and @uumesh for the suggestions leading to this PR! Hopefully I got your github handles right.