Runners use separate directories to allow parallelism

[Anders429]

This PR resolves the issues that are run into when running two separate unit tests in the same crate (see issue #58). Previously, doing this would cause the test runners to use the same crate directory, leading to one overwriting the other's manifest. This resulted in confusing errors and tests failing to no fault of the user.

To solve this, separate directories are defined for each run, using a hash of the test paths used by the TestCase. Therefore, the following now works correctly:

use trybuild::TestCases;

#[test]
fn test_1() {
    TestCases::new().compile_fail("compile_failures/1.rs");
}

#[test]
fn test_2() {
    TestCases::new().compile_fail("compile_failures/2.rs");
}

This does add a dependency on rustc-hash, the hasher used within rustc. I would say this is a trustworthy additional dependency, as it is owned by the Rust core team. The hash is not cryptographic, but it doesn't need to be for this use case. If you would rather not add this dependency, let me know and we can discuss another way.

[Anders429]

Ah, looks like rustc-hash would cause a bump in MSRV. After looking into it a little more, it looks like we can just use the DefaultHasher from the standard library, and it will accomplish the same thing without the extra dependency. I think that's the best way to go, since I don't want to unnecessarily cause bumps in the MSRV :)

[dtolnay]

I would prefer not to do this. In this approach the compilation of dependencies is not able to be shared between different trybuild invocations, and all the test outputs would still get interleaved/unreadable.

I would accept a PR to serialize the concurrent tests to run one after another using a file lock.

Thanks!