Better collapsing for assembler failures

[acfoltzer]

Since the error messages for llvm-as vary more than the ones from llvm-disasm, it's harder to write a good collapsing function for these. This means that almost all errors at that stage, many of which are dupes, move onto reduction (if enabled), which takes a very long time. We could probably get some mileage by comparing substrings in the error message to narrow down the bits that don't mention specific line numbers or functions from the assembly.

[langston-barrett]

I believe this issue is likely in reference to the fuzz-llvm-disasm tests.