For FP atomics involving bfloat vs. half, we need to distinguish the
type and not rely on the bitwidth alone. For my purposes, an alternative
would be to relax the atomic predicate MemoryVT pattern check with a memory
size only check. Since there are no extending operations involved,
the pattern value check should be unambiguous.

For some reason when using the _32 variants for atomicrmw fadd, I was able to
select v2f16 but v2bf16 would fail.

Changes mostly done with sed, e.g.
sed -E -i -r 's/atomic_load_(add|swap|sub|and|clr|or|xor|nand|min|max|umin|umax|swap)_([0-9]+)/atomic_load_\1_i\2/' llvm/lib/Target/*/*.td