feat(agent): add WebSocket connection layer

  Adds the Agent sub-client, AgentHandle, AgentEventStream, and
  AgentEvent types that connect to wss://agent.deepgram.com/v1/agent/converse
  and tie together the message and response types from prior commits.

  Surface:
    - Deepgram::agent() returns Agent<'_>; Agent::start() opens a session
      and returns (AgentHandle, AgentEventStream).
    - AgentHandle exposes send_settings/send_update_speak/send_update_think/
      send_update_prompt/send_inject_user_message/send_inject_agent_message/
      send_function_call_response for JSON messages, send_data for binary
      audio frames, and keep_alive/close for control. Method naming
      follows the existing FluxHandle convention (send_data for audio,
      plain verbs for control), with send_<message_type> for the typed
      JSON sends.
    - AgentEvent::Json(AgentResponse) | Audio(Bytes) is the unified
      stream item type — JSON events and audio frames are interleaved
      in their natural ordering. Matches the Python and JS SDKs.

  Connection details:
    - URL is a hardcoded constant (wss://agent.deepgram.com/v1/agent/converse).
      Self-hosted agent support is on the 0.10.0 roadmap; for now this
      is fixed.
    - Auth uses the Deepgram client's auth method (Token X / Bearer X).
    - dg-request-id from the upgrade response is parsed when present;
      falls back to the server's Welcome event when absent.

  Plumbing:
    - The agent feature now pulls in tungstenite + tokio-tungstenite
      (shared with listen).
    - WsError, TungsteniteError, and the From<TungsteniteError> impl are
      extended from cfg(feature = "listen") to
      cfg(any(feature = "listen", feature = "agent")) so agent works
      standalone.
    - auth field's allow-unused extended to cover agent; base_url and
      client stay listen-gated since agent doesn't use them today.

  128 tests pass (3 new), doctest in the module example compiles.
  Examples and an integration test against a mock server follow in
  subsequent commits.

Author: GregHolmes

Cargo.toml

@@ -59,7 +59,7 @@ default = ["manage", "listen", "speak"]
 manage = []
 listen = ["dep:tungstenite", "dep:tokio-tungstenite"]
 speak = []
-agent = []
+agent = ["dep:tungstenite", "dep:tokio-tungstenite"]
 
 [[example]]
 name = "grant_token"

src/agent/mod.rs

@@ -27,10 +27,14 @@
 //!   `SpeakUpdated`, `ThinkUpdated`, `InjectionRefused`,
 //!   `FunctionCallResponse`) plus an `Unknown` catch-all for forward
 //!   compatibility, all unified under the [`response::AgentResponse`] enum.
+//! - [`websocket`] — the [`Agent`] sub-client and live-session
+//!   primitives ([`AgentHandle`], [`AgentEventStream`], [`AgentEvent`])
+//!   that connect to `wss://agent.deepgram.com/v1/agent/converse`.
 //!
 //! Wire format matches the AsyncAPI schemas in `deepgram-docs` under
-//! `api/specs/asyncapi/schemas/agent/`. The WebSocket connection helpers
-//! (builder, handle, stream) land in subsequent commits.
+//! `api/specs/asyncapi/schemas/agent/`. Examples and additional
+//! convenience layers (e.g. file-based audio input) follow in
+//! subsequent commits.
 
 pub mod audio;
 pub mod aws_credentials;
@@ -42,6 +46,7 @@ pub mod response;
 pub mod settings;
 pub mod speak;
 pub mod think;
+pub mod websocket;
 
 pub use audio::{
     AudioConfig, AudioContainer, AudioInput, AudioInputEncoding, AudioOutput, AudioOutputEncoding,
@@ -77,3 +82,4 @@ pub use settings::{
 };
 pub use speak::{SpeakProvider, SpeakSettings};
 pub use think::{ContextLength, FunctionEndpoint, ThinkFunction, ThinkProvider, ThinkSettings};
+pub use websocket::{Agent, AgentEvent, AgentEventStream, AgentHandle};