Having a nested function allows us to discontinue the use of the `key`
variable which can be useful if multiple sections of unread_counts need
to be updated for a message type.

It includes type annotation changes such as:

* `key` passed into the nested function is typed using TypeVar.
* Message has been typed using a TypedDict so that mypy can determine
  the type of `key`. Note the use of Totality here because stream
  messages/pms/huddles do not have the same attributes. This change
  cannot be split into a prior commit because that will require
  multiple type definitions of the same `key` variable.
* An alternate declaration of `messages` in initial_index is necessary
  due to a mypy bug - python/mypy#7217 - while
  using TypeDict with defaultdict.

Note the use of Totality here because stream messages/pms/huddles
do not have the same attributes.

An alternate declaration of `messages` in initial_index is necessary
due to a mypy bug - python/mypy#7217 - while
using TypeDict with defaultdict.

Note the use of Totality here because stream messages/pms/huddles
do not have the same attributes.

An alternate declaration of `messages` in initial_index is necessary
due to a mypy bug - python/mypy#7217 - while
using TypeDict with defaultdict.

Note the use of Totality here because stream messages/pms/huddles
do not have the same attributes.

An alternate declaration of `messages` in initial_index is necessary
due to a mypy bug - python/mypy#7217 - while
using TypeDict with defaultdict.

Note the use of Totality here because stream messages/pms/huddles
do not have the same attributes.

An alternate declaration of `messages` in initial_index is necessary
due to a mypy bug - python/mypy#7217 - while
using TypeDict with defaultdict.

Note the absence of Totality here because stream messages/pms/huddles
do not have the same attributes.

An alternate declaration of `messages` in initial_index is necessary
due to a mypy bug - python/mypy#7217 - while
using TypeDict with defaultdict.

Fixes #3832
Fixes #5723
Fixes #17174
Improves #7217

This is a sixth "major" PR toward
#7724. Previously access to
"static" attributes (like type aliases, class objects) was duplicated in
four places:
* In `analyze_ref_expr()`
* In `determine_type_of_member()` (for modules as subtypes of protocols)
* In instance attribute access logic
* In class attribute logic

Most of these were somewhat incomplete and/or inconsistent, this PR
unifies all four (there is still tiny duplication because I decided to
limit the number of deferrals, i.e. preserve the existing logic in this
respect). Some notable things that are not pure refactoring:
* Previously we disabled access to type variables as class attributes.
This was inconsistent with plain references and instance attributes that
just return `Instance("typing.TypeVar")`.
* Instance access plugins were only applied on `TypeInfo`s and
`TypeAlias`es, now they are applied always.
* Previously arguments kinds were sometimes not correct for TypedDict
class objects with non-required keys.
* I tweaked `TypeOfAny` in couple places to be more logical.