Add announcement role (announcement@v1) skeleton

[marcelveldt]

Draft implementation skeleton for the announcement@v1 role proposed in Sendspin/spec#126 — short per-client audio clips (TTS, chimes, alerts) decoded on a second concurrent pipeline next to the media stream, with the ducking policy delivered to the embedder. Opening this alongside the spec PR so the discussion has a concrete client-side shape to poke at; expect it to track whatever the spec discussion decides (including the role id, if we stage it as announcement@_draft_r1 first).

What's here

• Wire surface: announcement@v1 role id, AnnouncementSupportObject (formats + own buffer_capacity) in client/hello, stream/start.announcement parsing (codec params + media_duck_db 0–50 / duck_ramp_ms 0–2000 / optional volume, out-of-range values dropped to defaults), client/state.announcement (state, required_lead_time_ms), binary type 24 (SENDSPIN_ROLE_ANNOUNCEMENT = 6, routed through the existing get_binary_role() helper unchanged).
• AnnouncementRole following the documented add-a-role pattern (public header + pimpl + inbox slot + shared event ring), with an AnnouncementRoleListener the embedder implements: on_announcement_write() (PCM, task thread) plus on_announcement_start/end/clear() (main loop). on_announcement_start() carries the duck parameters — that's where e.g. ESPHome calls mixer_speaker.apply_ducking(); on_announcement_end() fires on server end, local completion, abort, and transport loss alike, so media can never stay stuck ducked.
• AnnouncementTask: a dedicated lightweight decode task (own ring buffer + SendspinDecoder instance, one extra thread, default priority below the media sync task). Deliberately no Kalman/sync machinery: announcements are per-client, so the task starts output at the first chunk's timestamp via the time filter (or immediately without sync) — honoring that timestamp is also what aligns a coordinated multi-speaker announcement — never drops late chunks, is paced by the sink's blocking writes, and ends a stalled stream locally after 5 s.
• Lifecycle wiring in SendspinClient: stream/end with omitted roles ends announcements too; stream/clear with omitted roles (a media seek) explicitly does not touch the announcement stream — only an explicit roles: ["announcement"] does.
• Build gates: SENDSPIN_ENABLE_ANNOUNCEMENT (Kconfig default n while the spec is draft; depends on SENDSPIN_ENABLE_PLAYER since it reuses the ring buffer + decoder sources).
• Tests: protocol round-trip coverage for the hello support object, stream/start defaults/validation, client/state serialization, and the binary allocation. Host build compiles clean (-Wall -Wextra -Wpedantic), full suite passes.

Deliberately not here (yet)

• The ESPHome layer (a SendspinAnnouncementSource feeding the speaker_source announcement pipeline + duck-parameter plumbing) lives in the esphome repo and needs a released sendspin-cpp first.
• Gain handling: per the spec draft, duck and announcement volume are applied by the embedder at the sink/mixer (where media volume already lives) — the library only transports the parameters.
• The stream-replace (stream/clear) boundary is approximate in the task (buffered chunks racing the drain may go with the old clip); fine for silence-padded replacement clips, noted inline.
• Announcement-capable secondary connections (spec open question 6 — e.g. HA announcing while another server streams) would live in the connection manager / admission logic, not in this role.

Marked draft until the spec PR settles.