Tests with signatures, fixes #1074

[protolambda]

Made the pytests respect BLS signatures. This enables us to use the tests to generate properly BLS-signed test-vectors with.
Note: there are a few exceptions:

• finality tests are too slow (many transitions), hence they run with disabled BLS. I'm not sure how/when we include these in the test-vectors for clients
• there are a few sanity tests that I just disabled, as they cover long-range (empty slots) block transitions, costing lots of hashing.

The block-operations and epoch transitions are fully signed now 🎉

TODO:

• Invalid-signature tests for block operations
• Cover more of epoch transition with new tests future PR
• Call the tests in generator-mode to generate test-vectors. (epoch transition generator in future PR)
• Implement some basic SSZ caching, to speed up tests (the hashes of fields of empty blocks being cached would help a lot! Tests typically change only a part of the block). CI runs fast enough, SSZ caching to be implemented in separate PR

[djrtwo]

Ah, looks like this branch doesn't have the changes from #1052 integrated. Is the intention to modify generate_from_tests to override the default of BLS being off to on?

[hwwhww]

IMO it's safer to make it more explicit by removing the default value False from signed parameter.

[protolambda]

@djrtwo

Ah, looks like this branch doesn't have the changes from #1052 integrated.

It's branched off of that, so it should. You mean the proposal for BLS test-improvements? This PR is in draft mode mostly because:

• want to implement some invalid-signature tests
• need to go over the BLS things listed in this comment: BLS and testing #1074 (comment) to see what we really want, and what is better to adjust. (i.e. test-markers for BLS set to always ON)

Is the intention to modify generate_from_tests to override the default of BLS being off to on?

Ideally I think BLS should be ON, if pytests were fast enough. But we already discussed this, and it will be off for CI where possible, but ON where needed + in generated vectors.

For now I just like the explicitness, as it helped me tracking what needs to be updated, and finding testing bugs (things not being signed). If things are stable we may go for implicit signed True/False, but I don't think it's a netto win now.

[djrtwo]

Last set of cleanups and forced-on bls tests look good.

Headed to bed. good to merge when you're up