docs: custom reporting tags the Voice API - #421

Tagging a Voice session is confirmed working in production, but the docs
still said custom tag data was supported for text translation only.

- Rework the how-to guide: generalize the framing away from "translate
  requests" and add a section on tagging a Voice session, including that
  the header can only ride the session request (the WebSocket carries no
  headers), so a tag covers a whole session and cannot be changed
  mid-session or applied per utterance.
- Add speech_to_text_minutes and speech_to_speech_minutes to
  CustomTagBreakdown, the analytics response examples, and the endpoint
  reference table, and note that voice minutes are not part of
  total_characters.
- Document X-DeepL-Reporting-Tag in the spec as a reusable
  components.parameters entry, referenced from POST /v2/translate and
  POST /v3/voice/realtime. The header was previously absent from the
  spec entirely, including for text translation.
- Add an August 18 release note and bump info.version to 3.13.0.

openapi.json is left untouched; CI regenerates it from the YAML.

- Add the missing sample response to the Voice curl example, since
  CLAUDE.md requires request and response together and the translate
  section directly above it already shows one.
- Replace an em-dash (banned) and drop "utterances", which appears
  nowhere else in the docs, in favour of "audio chunks" as used
  throughout the Voice pages.
- Restore "only" in both scope warnings. Dropping it made the constraint
  read weaker than it is.
- Fix the total_characters description, in the reference table and in the
  spec, which still claimed all services while voice minutes now sit in
  the same breakdown and are excluded from that total.
- Note that the minute fields return fractional values.
- Lead the release note with what a custom tag is, per the changelog
  convention, so the entry stands alone.
- Trim the CustomReportingTag description down to purpose plus a link to
  the guide, in line with sibling parameters.
- Say "text translation" throughout rather than alternating with
  "translate", and cross-link the tag from the voice session flow and
  from the guide's own Limitations list.

Verified against production: a tagged session streamed 8800 ms, dropped
the WebSocket without ending the stream, reconnected via
GET /v3/voice/realtime with only the token (no tag header), then streamed
8754 ms more. All 17553 ms landed under the original tag across four
usage rows, with no untagged rows for the account.

This was the gap flagged in review. Reconnection is a normal flow and the
reconnect request cannot carry a tag, so silence here left developers
unable to tell whether reconnected audio kept its attribution.