Replace the hand-rolled schemaless_struct_type() with a compile-time
schema descriptor system. The schema is defined as a type:

```
  using schemaless_desc = struct_desc<
    field_desc<"redpanda", rp_desc>>;
```

From this single definition:
- schemaless_struct_type() generates the runtime iceberg::struct_type
- index_of<"field_name"> provides compile-time field indices
- type_field/value_field accessors replace brittle fields[0]
- rp_struct_type()/rp_struct_value() provide named access

So code like this:

```
std::unique_ptr<iceberg::struct_value> build_rp_struct(
  model::partition_id pid,
  kafka::offset o,
  std::optional<iobuf> key,
  model::timestamp ts,
  model::timestamp_type ts_t,
  const chunked_vector<model::record_header>& headers) {
    auto system_data = std::make_unique<iceberg::struct_value>();
    system_data->fields.reserve(6);

    system_data->fields.emplace_back(iceberg::int_value(pid));
    system_data->fields.emplace_back(iceberg::long_value(o));
    // NOTE: Kafka uses milliseconds, Iceberg uses microseconds.
    system_data->fields.emplace_back(
      iceberg::timestamptz_value(ts.value() * 1000));
    ... brittle, carefully ordered code ...
```

...becomes:

```
std::unique_ptr<iceberg::struct_value> build_rp_struct(
  model::partition_id pid,
  kafka::offset o,
  std::optional<iobuf> key,
  model::timestamp ts,
  model::timestamp_type ts_t,
  const chunked_vector<model::record_header>& headers) {
    return rp_desc::build_value(
      val<"partition">(iceberg::int_value(pid)),
      val<"offset">(iceberg::long_value(o)),
      val<"timestamp">(iceberg::timestamptz_value(ts.value() * 1000)),
```

and the definitions of compile-time schemas becomes more crisply defined
as well.