Performance regression in phase averaging

[kbarros]

From @Lazersmoke:

https://github.com/SunnySuite/Sunny.jl/commit/3c43a43685fac5a86e30a99f4dc1d4bf1baf8521 reintroduced a lot of allocations into the inner loop that actually slow things down, and partially reverting (keeping the change to q instead of q_absolute) made it fast enough for my code to run again.

Strangely, when I committed that same patch, I benchmarked a very large speedup. Test files attached. MgCr2O4-test.zip

[kbarros]

Should now be fixed. For this updated benchmark, MgCr2O4-test.zip, I measure:

• About 8s on Sunny v0.5.8.
• 1.63s after reverting https://github.com/SunnySuite/Sunny.jl/commit/3c43a43685fac5a86e30a99f4dc1d4bf1baf8521
• 1.27s after keeping https://github.com/SunnySuite/Sunny.jl/commit/3c43a43685fac5a86e30a99f4dc1d4bf1baf8521, and additionally applying the fix https://github.com/SunnySuite/Sunny.jl/commit/90b9480c5c329a5f1023dd5677cd77bf56099474.

That is, Sunny 0.5.9 will be 30% faster than any previous Sunny. The fix was to pass length information to ntuple statically, which avoids reliance on compiler inlining.