Better error diagnostic for matmul, where inputs and result are all bf16.

[newling]

This matmul is not supported, because we expect accumulation to be in float32. It should be supported though, with some casting