Previously errors were sorted by Symbol index instead of the string. The
indexes are not the same between architectures because Symbols for
architecture extensions (e.g. x86 AVX or RISC-V d) are interned before
the source file is parsed. RISC-V's naming of extensions after single
letters led to it having errors sorted differently for test cases using
single letter variable names. Instead sort the errors by the Symbol
string so that it is stable across architectures.
