Fix morph negation transforms

[SingleAccretion]

The short version of the bug is that the transforms in question did not account for persistent side effects, which could lead to silent bad codegen. Also, the GTF_EXCEPT guards were unnecessary for transforms that did not reorder operands.

This change leads to one diff in the benchmarks, a regression of 2 bytes, caused by the fact that gtCanSwapOrder does not understand that we can swap when both trees have GTF_GLOB_REF set, but no other effects on them:

-               [000078] ----G+------              \--*  SUB       int
-               [000077] ----G+------                 +--*  LCL_VAR   int   (AX) V11 loc6
-               [000075] ----G+------                 \--*  LCL_VAR   int   (AX) V08 loc3
+               [000078] ----G+------              \--*  ADD       int
+               [000076] ----G+------                 +--*  NEG       int
+               [000075] ----G+------                 |  \--*  LCL_VAR   int   (AX) V08 loc3
+               [000077] ----G+------                 \--*  LCL_VAR   int   (AX) V11 loc6

Fixes #55140.

Note: I am aware that gtCanSwapOrder is not without its problems as well, but it still seems decidedly better to use it instead of bespoke guards.

[jakobbotsch]

Thanks for fixing this!