This reverts commit be17791f2624f22b3ed24a2539406164a379125d.

This is not necessary for gfx1250 anymore.

This would select the pseudo and then crash when the MC instruction
was used. I believe this has been broken since 9912ccb0b4d17

Re-enable it by adding a wait on vm_vsrc before every barrier "start"
instruction in GFX10/11/12 CU mode.

This is a less strong wait than what we do without BackOffBarrier, thus
this shouldn't introduce
any new guarantees that can be abused, instead it relaxes the guarantees
we have now to the bare
minimum needed to support the behavior users want (fence release +
barrier works).

There is an exact memory model in the works which will be documented
separately.

This now applies to gfx90a+ and gfx1250, so add a dedicated feature
for it so the feature check can be composed correctly with unrelated
properties. Probably would be better to invert this, but that turns
out to be difficult given the current assumptions.

It is the same as reading SHADER_CYCLES_LO and SHADER_CYCLES_HI
but with a single instruction.

HW will emulate unsupported PCIe atomics via CAS loop, we do not need to
expand these anymore.