Handle positive and negative infinity in TopKV2.

TopKV2 hides iota in the low bits of the input after converting from bf16 to f32. This usually works, but for positive and negative infinity or'ing in iota produces NANs.

To handle positive and negative infinity, treat bf16 as integers in
sign-magnitude format. Convert to two's complement. Sort in two's complement and
convert back.

Add an exhaustive unit test for bfloat16 to float conversion.

PiperOrigin-RevId: 201421784