…nz-floss#285)

`pagination_info_from_proto()` calls the `PaginationInfo` constructor
directly, so the `ValueError` that frequenz-floss#217 added for a negative
`total_items` can escape from a conversion function. That breaks the
wrapping rule that loading from the wire never raises on invalid
content: the caller loses the whole message over one bad field.

This adds the usual valid/invalid pair and the replacement converter:

* `BasePaginationInfo` holds the two shared fields and refuses direct
instantiation.
* `PaginationInfo` keeps its `ValueError` unchanged. There is
deliberately no `_raise_on_invalid`-style escape hatch: unlike
`DeliveryArea`, it has always been strict here, so there is no released
behavior to transition away from.
* `InvalidPaginationInfo` takes the same fields with no invariants, so a
malformed count survives for inspection.
* `pagination_info_from_proto2()` returns `PaginationInfo |
InvalidPaginationInfo`; `pagination_info_from_proto()` is deprecated and
points at it, mirroring `delivery_area_from_proto` and
`bounds_from_proto`.

Two things worth a second look while reviewing:

The new converter reads `next_page_token` through `HasField()` instead
of testing it for truth. The field has presence, so the old truthiness
test collapses "the server sent an empty token" into "the server sent no
token", and only the second means there are no more pages. Changing that
in place would be a silent behavior change, so it only happens under the
new name.

`total_items` is a `uint32` on the wire, so a decoded message cannot
actually carry a negative count today and the invalid branch is
unreachable through the generated bindings. It is still there because
the converter's contract is not to raise regardless of what it is
handed, and because the wire type is not the wrapper's to assume; the
test exercises it with a message-like stand-in.

Both leaves also gain a `__str__`, which neither had. The pair only
reads well if the invalid marker composes with a valid form, and the
dataclass `repr` is too long for a log line. The rule is per-field, so
the marker goes on the count (`items=<invalid:-1>,next=None`) rather
than around the whole object, as in `InvalidDeliveryArea`.

The release notes document the deprecation and, separately, the breaking
change from frequenz-floss#217: `PaginationInfo(total_items=-1)` raised nothing in
v0.4.0 and raises `ValueError` now.

Part of frequenz-floss#239 and frequenz-floss#251.

Fill the gaps found in a pre-release sanity check of `RELEASE_NOTES.md`:

* Replace the template placeholder with a real release summary,
including release stats and the note that this is a huge release despite
the patch version, realizing the new library design while keeping
backwards compatibility, with the breaking cleanup to follow in v0.5.0.
* Document two breaking changes to symbols released in v0.4.0 that were
missing from Upgrading: `Bounds` raising on `NaN` endpoints (frequenz-floss#254) and
`MetricConnection.name` no longer being optional (frequenz-floss#257).
* Complete the `UNSPECIFIED`-as-`0` bullet with `metric_config_bounds`
keys (frequenz-floss#224), merge the duplicated `Microgrid` bullets, and briefly cover
the `ElectricalComponent` accessors, `CategorySpecificInfo` and the
smaller `Invalid*` types from frequenz-floss#237, frequenz-floss#248, frequenz-floss#249, frequenz-floss#257, frequenz-floss#268 and frequenz-floss#273.
* Add the `str()` exception message change (frequenz-floss#258) to Bug Fixes.

The `PaginationInfo` breaking change from frequenz-floss#217 is documented separately
in frequenz-floss#285, which adds `InvalidPaginationInfo` and
`pagination_info_from_proto2`.