Optimized dot and iamax routines

[tdehoff]

Significantly improves the performance of dot and iamax routines.
Tested on float32, n = 134217728:
iamax was 5.24083 GBs, now 1032.5771.
dot was 631.325 GBs, now 1418.0388.

[tdehoff]

nrm2 was 380.57462 GBs, now 1104.4181 GBs.
It looks like you can't do sqrt on float64s on NVIDIA, so I cast to Float32 before applying sqrt. Surprisingly, it still passes correctness tests on my side.

[jacksonmowry]

I'm seeing ~2% of dot tests fail when I run 500 tests on both the main branch and for this PR. So nothing got worse, but something we should probably look into in the future. I'll create an issue for it.

Other than that looks good!

[tdehoff]

@jacksonmowry it looks like BLAS++ uses check_gemm function to test dot error: https://github.com/icl-utk-edu/blaspp/blob/master/test/test_dot_device.cc. I'll fix our tests and see if this helps.