Runtime benchmark codegen diff

[Kobzol]

This PR adds a new command to collector, called codegen_diff. It takes as input the name of a single runtime benchmark group, the type of codegen you are interested in (assembly, assembly interleraved with Rust source, LLVM IR or MIR) and two versions of the compiler. It then uses the cargo-show-asm tool to compute the codegen and diff it between the two compiler versions.

The diff is computed separately for each symbol (function) found in the benchmark crate. We therefore have to match the same function between two rustc versions (toolchains) by its symbol name. By default, cargo asm doesn't include hashes in the symbol names, which helps making sure that the name will stay the same, but it also means that a different version of the same function (e.g. one monomorphized with a different generic argument) will get matched between two toolchains. There is a cargo asm flag called --full-name which includes the hashes in the symbol names, which would resolve this issue. However, there is a problem with that - different versions of the compiler produce different hashes, and this doesn't seem to be easily fixable. Therefore, I don't include the hashes in the symbol names, and I also opted in to the v0 mangling scheme, as it seems to me that with it more functions are matched even without the hashes.

The diff for each function is then printed to stdout, with colors if the output is a terminal. Eventually we might want to display some small HTML page so that the diff is easier to understand (?).

[Kobzol]

Thanks!