[llvm] [AMDGPU][GlobalISel] Legalize odd bf16 vectors (PR #226707)

[Gheorghe-Teodor Bercea via llvm-commits]

https://github.com/doru1004 approved this pull request.

LGTM with nits:
- The four new strict_*_v3bf16 tests have no `#0`
- each new v3 test in fadd, fma, fminnum, fmul and fsub runs straight into the existing v4 test without leaving the appropriate blank lines in between.

https://github.com/llvm/llvm-project/pull/226707