MSL: divergence in different code stages when relying on DecorationInvariant

[mabaro]

The current implementation relies on adding the dependency chain of the invariant variable (i.e, position) to the forced_temporary group. This already splits a lot of operations reducing the chances of divergencies, but it's not enough for complex setups.
In our use case, the spirv-opt output has divergencies on different stages, splitting one particular operation on one stage (i.e., 961 - (_958 * dot(_961, _958))) into temporaries, while keeping it inlined on the other stage.
When passing the produced spirv through spirv-cross the invariant decoration current behaviour keeps the divergences generating glitches.
The only solution I came up to have some sort of working code is:

• propagate decorationInvariant through the dependency chain
• use decorationInvariant in the same sense as decorationNoContraction in emit_binary_op/emit_binary_op_func (i.e., the latter is used by spirv_msl when outputting spvXXX functions)

This is for sure not a complete solution, and I don't know if that approach could even be considered adequated for merging at certain point, or for further investigation.

I'd appreciate any hints/opinions.
Cheers.

[mabaro]

related post -> #1022