[TF:XLA] do not emit bfloat16 sum reductions from tf2xla

[TF:XLA] do not emit bfloat16 sum reductions from tf2xla

bfloat16 is a storage format, not a computation format. Doing reductions in
this reduced precision is prone to quickly overflow.  Instead, emit a float32
computation, and wrap the reduce params and result in conversions to and from float32.

PiperOrigin-RevId: 189977590