[llvm] [SelectionDAG] Avoid double rounding by using a libcall for int (>i32) to bf16 conversions (PR #71658)

[Eli Friedman via llvm-commits]

efriedma-quic wrote:

As far as I can tell, DAGTypeLegalizer::SoftPromoteHalfRes_XINT_TO_FP is still doing the same broken conversion it always has.

https://github.com/llvm/llvm-project/pull/71658