We only ever need to upcast to shared references.

This change isn't necessary, just dead code
cleanup.

Also, move to a blanket impl'd trait.
Overall cleaner approach.

The traits are now quite simple:

* Database is the external trait
* ZalsaDatabase is the internal one, implemented
  by `#[salsa::db]`. It adds two methods,
  `zalsa` and `zalsa_mut`. Those give access
  to our internal methods.

For now I've hidden the methods behind
`&dyn Zalsa`. This is nice and clean but it may
be worth later refactoring to a `struct Zalsa`.

The distinction is dumb and should go away.
But we still need it for a bit.

Creating events if nobody is listening has
always bugged me.

Under this design, *all* databases are a
`DatabaseImpl<U>`, where the `U` implements
`UserData` (you can use `()` if there is none).

Code would default to `&dyn salsa::Database` but
if you want to give access to the userdata, you
can define a custom database trait
`MyDatabase: salsa::Databse` so long as you

* annotate `MyDatabase` trait definition of
  impls of `MyDatabase` with `#[salsa::db]`
* implement `MyDatabase` for `DatabaseImpl<U>`
  where `U` is your userdata (this could be a
  blanket impl, if you don't know the precise
  userdata type).

The `tests/common/mod.rs` shows the pattern.

Separate handles are no longer needed.

Each clone gets an independent local state.

The aim is to eventually eliminate Runtime.

This should really be synchronized with the
codespaces and github configuration but...
I'm not clever enough to do all that.

and add more comments

Co-authored-by: Micha Reiser <[email protected]>

Fixes salsa-rs#520