Updating the CI to build different precisions

[Ricardicus]

Adding the BF16, FP16 and FP32 builds to the Github workflow.
I am also passing the defines to the testing and profiling build.

[Ricardicus]

It took some attempts before I got it to build... Turns out I needed to raise the cuda image version for the bf16 to compile. Maybe a change in the bf16 api had been introduced.

There is another issue I don't know much about that breaks the build. But I think this PR is kind of solid now.

[karpathy]

ok ty, though i'm noticing the CI is failing here? hmm

[Ricardicus]

I think you fixed the CI, it was this error:

"
Run OMP_NUM_THREADS=8 ./test_gpt2
Bad version in model file
Error: Process completed with exit code 1.
"

I ran it some time ago, the fix wasn't in back then.

[Ricardicus]

Let's see if it builds now.... I pushed and then rebased just to trigger a new build.

[Ricardicus]

Looking good 😎

[rosslwheeler]

@Ricardicus - this might already be in the code somewhere but when we run - is there a run-time way (a printf?) where we can tell which precision it was built with?