Add a new stress benchmark for including huge binary blobs with `include_bytes!`

[petrochenkov]

Around 10-20MB of data should be enough, I guess.

The compiler should do as little processing on such blob data as possible.
cc rust-lang/rust#65818

There's no such benchmark currently, so changes like rust-lang/rust#141875 (in its current version) will go unnoticed.

[Kobzol]

Yeah, that's a good point, I was thinking about that the last time I was taking a look at optimizing the include macros (IIRC there's still some overhead caused by inefficient LLVM hashing on strings). I'll take a look at this next week.

[Kobzol]

Although the "large data" part might be a bit problematic, because it means that we would have to include a 20 MiB file into the repository 🤔 But I guess if it's easily compressible text, it might not be so bad...

[petrochenkov]

because it means that we would have to include a 20 MiB file into the repository

Or a small script generating that file using Fibonacci or something (just to avoid randomness in benchmarks).

[nnethercote]

I think including include_str! as well would be good. (We could include the same file twice, once withinclude_bytesand once with include_str!.)

And maybe include! too, though I'm less sure about that.

[petrochenkov]

include!("foo.rs") is basically indiscernible from regular mod foo;, so not very useful.

[Kobzol]

#2175 adds a benchmark for this.