Follow-up for canonical se/de after an additional round of review

[xgreenx]

• Removed all enum attributes. The manual implementation of the traits can do the same.
• Removed clone for each Input required to call into_full. It is possible because of a new Empty<Type> type that implements canonical se/de with minimal implementation.
• Removed SIZE_NO_DYNAMIC because it is not used in the codebase.
• Removed SerializedSizeFixed because we don't have any specific logic that uses it.
• Moved all methods from SerializedSize to Serialize. It looks unnatural that #[derive(Serialize)] implements 3 traits. Instead, it implements on trait now.
• Removed SIZE_STATIC because it is impossible to calculate in the case of enums(or if the struct has an enum field). Supporting something like this requires another way of doing that, plus it doesn't provide a lot of wins for us right now because #[inline(always] handle that well enough.
• Removed panics during size calculation. Instead we use saturating_add. If the size is incorrect we will fail in another place.
• Removed usage of num_enum and replaced with canonical se/de.
• Removed calculation of the discriminant on macro level. Replaced with native rust calculation that covers all cases.

[Voxelot]

What's the benefit of implementing this kind of boilerplate logic manually? Does it just add unnecessary complexity to the macros by trying to automate this?

[xgreenx]

We can use the peek function, which removes the Clone requirement from the trait Input: Clone. Skipping of the discriminant is a very "magic" thing and I prefer to do that manually because we need it only in one place for Transaction.

[Dentosal]

Removed all enum attributes. The manual implementation of the traits can do the same.

I dislike the added copy-paste, as that makes it possible to accidentally apply a change to only a part of the branches.

Removed clone for each Input required to call into_full. It is possible because of a new Empty type that implements canonical se/de with minimal implementation.

The empty type changes I really like, and this is a nice additional benefit.

Removed SIZE_NO_DYNAMIC because it is not used in the codebase.
Removed SerializedSizeFixed because we don't have any specific logic that uses it.
Moved all methods from SerializedSize to Serialize. It looks unnatural that #[derive(Serialize)] implements 3 traits. Instead, it implements on trait now.

Sure.

Removed SIZE_STATIC because it is impossible to calculate in the case of enums(or if the struct has an enum field). Supporting something like this requires another way of doing that, plus it doesn't provide a lot of wins for us right now because #[inline(always] handle that well enough.

I though there was an attempt to get const-time size calculations. If this is no longer a goal, then removing the const-time size access is a good thing.

Removed panics during size calculation. Instead we use saturating_add. If the size is incorrect we will fail in another place.

I dislike this, as it has a possibility to silently fail, and could hide implementation bugs. Shouldn't matter much, but if we change the serialization itself at some point, this makes it a bit harder.

Removed usage of num_enum and replaced with canonical se/de.

Sure.

Removed calculation of the discriminant on macro level. Replaced with native rust calculation that covers all cases.

The new code seems good enough.

[Dentosal]

I disagree with some design decisions here, as mentioned above.
Still, I don't think it matters enough to keep discussing this, so lets just go ahead and merge.