Persist binary media for session suspend and resume

[Sewer56]

Status

Not planned. Unless I start using with TUI/GUI/non-headless.

We cannot currently suspend and resume sessions (across process boundaries) because the framework discards all binary media after a run completes.

This issue proposes an opt-in way to persist and retrieve this media, so a resume
can happen.

Motivations

A run's binary media exists only for the run's lifetime:

• Distillation redacts user-to-LLM inline binary (images, audio,
  video, documents, files) to a media type plus byte count in audit
  text, then drops it.
• Distillation skips LLM-to-user binary (ModelResponsePart::File)
  entirely. The caller can neither observe nor retrieve it.

Zero copies survive RunComplete processing. That is deliberate for
the audit trail, but it makes suspend and resume impossible:

• Resume needs the raw transcript, including binary, to rebuild the
  provider request prefix.
• RunConfig::preamble_messages cannot substitute: it flattens text
  into the prompt string, losing tool-call structure and media.

A session that stops loses every image it ever held.

Solution

Opt-in session persistence:

1. Capture: an opt-in way to obtain the raw vendor transcript
   (e.g. serdesAI's Vec<ModelRequest>).
2. Persist: serialize that transcript to caller-chosen storage.
   Binary survives the round trip.
3. Resume: start a new run seeded with the persisted transcript via
   the vendor's message_history.
4. Retrieve: a way for the caller to read LLM-generated
   ModelResponsePart::File content instead of losing it.

The caller decides where files live (single session file, or a
content-addressed blob store). etc.

Acceptance criteria

• A persisted-then-resumed run reproduces the original serialized
  request prefix, so provider prompt caches still hit.
• Binary user parts resume with full fidelity: the resumed model sees
  the same media bytes.
• The caller can retrieve LLM-generated ModelResponsePart::File
  content.
• Persistence is opt-in; runs without it keep the current
  zero-copy-after-completion behavior.
• Resumed runs do not duplicate the static system prompt.

Additional context

Memory footprint is a stated project selling point (~10 MiB, README).
Persisted transcripts live on disk, not in memory, so the default
path stays unchanged.