tests: Make known_failure more precise

[moulins]

(Requires PR #22348)

Instead of treating any error or panic as a success, only "expected" failures are allowed, and the kind of check that should fail must be specified:

• root-level known_failure = true: means that the final trace output check is expected to fail;
• root-level known_failure.panic = "msg": means that the test is expected to panic (like the #[should_panic = "msg"] attribute in Rust unit tests);
• image_comparisons.NAME.known_failure = true: means that the given image check is expected to fail.

[Dinnerbone]

Tests in --release now fail (I do this a lot, it's faster to test in release 😅)

It's because we expect on e.g. overflow panic message, but those don't exist in release.

        FAIL [   0.234s] tests::tests [!] avm1/bitmap_data_thorough/pixelDissolve
        FAIL [   0.067s] tests::tests [!] from_gnash/misc-ming.all/get_frame_number_test
        FAIL [   0.075s] tests::tests [!fp10] from_gnash/misc-swfc.all/matrix_accuracy_test1
        FAIL [   0.057s] tests::tests [!fp9] from_gnash/misc-swfc.all/matrix_accuracy_test1

[moulins]

Added a commit to skip panicky tests if debug assertions are disabled; it's a little heavy-handed, as some panicky tests hit regular assertions, but we have few enough of those that I don't think it's a big deal.