dividing a float16 tensor by a python float is inaccurate with nvfuser

[beverlylytle]

The sample case of test_core_vs_torch_consistency_div_nvfuser_cuda_thunder.dtypes.float16 which uses a python float as the denominator fails with numerical inaccuracy. If the same sample is run with a tensor with empty shape, the test passes. Specifically, if .item() is deleted from this line, the test passes.

cc @tfogal

[mruberry]

@beverlylytle, thank you for filing this issue. I know you're not suggesting otherwise, but if anyone else is reading this, it's OK to have a sampleinput that tests fp16 tensor / pyfloat, because that's a real supported input to div. We should figure out what's going on and see if we can fix it (or maybe just widen the test's tolerances)