move session handler traits to frame-support

[kianenigma]

Will help with paritytech/statemint#43.

Context

For a simple parachain like statemint, we are not going to use staking or session, as they are too complicated. But we still need to use pallet-aura. The only easy and sensible way to communicate with pallet-aura is to register it in the list of SessionHandler somewhere. This is exactly how relay chains work.

Now, I want a new pallet like parachain-selection to act like the combination of staking and session for a parachain, and select some collators in some simple way every N block and report them to pallet-aura.

All of this works, but it has two flaws:

1. SessionHandler trait was defined in pallet-session and in order to reuse it I had to import it, which was not good.
2. I can imagine why SessionHandler was defined in pallet-session: its API is exposing queued_validators which is specific to pallet-session. And what I have done now is breaking this. I am not happy with this but in short term I think it is okay and I will make a follow up issue with further details. Essentially, we need to convert SessionHandler trait to something more generic, so that other chains can write their own custom pallet-session, but still be able to use our aura (and grandpa, babe, authority-discovery pallets). This will need to also change the OneSessionHandler.

If it makes anyone happier, I can think of something like this, we could make the queued_validator an Option and that's already a good step, but tbh I felt like it doesn't make a big difference at this point and a big revamp is needed.

---

I also removed a bunch of useless TestSessionHandlers here and there.

[gui1117]

If it makes anyone happier, I can think of something like this, we could make the queued_validator an Option and that's already a good step, but tbh I felt like it doesn't make a big difference at this point and a big revamp is needed.

It would be great for Babe to ensure at compile-time that at least one session is queued.
Maybe we can create different trait depending on the number of session queued.

We could create trait SessionHandler and SessionHandlerWithQueued, aura would implement both, babe only the latter.

Or maybe we can solve it with some const like:

trait SessionHandler<const QUEUED: usize> { .. }

impl<T: SessionHandler<0>> SessionHandler<1> for T { .. }
impl<T: SessionHandler<1>> SessionHandler<2> for T { .. }

struct Babe;
impl SessionHandler<1> for Babe { .. }

[gui1117]

Now, I want a new pallet like parachain-selection to act like the combination of staking and session for a parachain, and select some collators in some simple way every N block and report them to pallet-aura.

moving the trait to support won't solve the fact that the trait is enforcing queueing 1 session. And if the session is queued by one then I don't see a need to implement it differently than pallet-session.
Why not replacing only staking by another pallet and keeping the use pallet-session in parachain ?
A pallet implementing SessionManager could solve the usecase no ?

I'm not opposed to move it to support. But I feel like if we are going to queue the sessions by 1 then pallet-session should fit our need and only another SessionManager implementation is needed.

[kianenigma]

A pallet implementing SessionManager could solve the usecase no ?

I like this idea as well, but I have to double-check. I was worried that pallet-session would be too much for a parachain and bring in too much functionality.

[kianenigma]

@thiolliere in that case I will rework this PR to move SesseionManager out of pallet-session to avoid the dependency, and I agree that is slightly better. SesseionManager is a more generic trait and exposes less of the underlying session pallet complexity.

[kianenigma]

not a big deal for now