Add lit tests; shows initial bfloat16 support for tt.dot

[abrown]

The "expand to target type" logic added in #231 works for bfloat16 types, not just float16. This change adds lit tests to show the general form of the lowering our convert-triton-cpu-to-llvm does when run.

[abrown]

I will note that performance for bfloat16 is quite bad compared to float16 (slower by an order of magnitude): e.g., when bfloat16 support is patched in to the 03-matrix-multiplication.py tutorial, I observe that each bfloat16 value is individually converted using __truncsfbf2 which delegates to some expensive logic (float2bfloat).

[alexbaden]

These look fine, though once #254 lands I imagine you would have to update the fmul and fadd bits. Might be easier to just check for promotion and assume the promoted value is used (best yet canonicalize in the RUN command so an unused fp extension gets dropped) .

[abrown]

I tried to additionally add:

    // COM: And at least 8 multiplications:
    // CHECK-COUNT-8: llvm.fmul

But I think they have to be sequential? I guess I can figure that out in a follow-on PR.

[alexbaden]

I'm fine with this for now with the caveats that long term (1) test_core.py::test_dot is preferred for correctness and (2) we should strive to make lit tests as focused as possible (e.g. only test promotion to fp32 during the llvm lowering).