Make generic composition information more efficient (#85623)

Make generic composition information more efficient (#85623)

Before this PR, we tracked information about generic composition of a `MethodTable` this way:

* Each generic instance `MethodTable` had a pointer to the generic definition `MethodTable`.
* Each generic instance `MethodTable` had a pointer to its generic composition: a variable sized data structure that had a small header (number of elements, a flag whether it's variant), followed by pointers to `MethodTable`s of components, optionally followed by variance information for each argument.

This works, but it's not particularly efficient, especially in light of some facts that didn't exist at the time this scheme was introduced.

In particular:

* The number of generic parameters can be obtained from the generic definition, no need to duplicate it into instances.
* The variance information can be obtained from the generic definition, no need to duplicate it into instances.
* It makes no sense to indirect to a single-element list - the list can be bypassed if arity is 1.

This PR addresses all of the above.

Saves about 0.5% in size for BasicMinimalApi, improves startup (the composition no longer needs to be dehydrated because it's relative pointers now), and improves working set (the composition stuff accounted for 100 kB of private dehydrated working set in BasicMinimalApi).