rocr: fixing wave_end trap handling

[rocm-devops]

1. Bug: Branching should be to .not_wave_start, not to .not_wave_end.
2. ttmp13 contains a copy of TRAP_CTRL and, according to gfx1250 SPG:

Unlike other trap causes, when the trap handler is entered because of trap-on-wavestart or trap-on-
waveend, the associated TRAP_CTRL bit is cleared to zero.

so, we should not be checking SQ_WAVE_TRAP_CTRL_WAVE_END_SHIFT, it will be always zero.

[rocm-devops]

Might be worth having the comments you had in the PR's message also present in the commit messages (looks like the policy is rebase and merge, which will preserve the commit messages). If this is a plain merge, one might want to make sure the "gfx125" part of the source branch does not end-up in the final tree.

Otherwise the changes themselves look good to me.

[jayhawk-commits]

This pull request has merge conflicts and cannot be automatically imported to ROCm/rocm-systems.