moq-net: auto-create Origin on connect/accept, expose via Session

[kixelated]

Summary

Makes explicitly constructing an Origin optional on both client and server in moq-net and moq-ffi. The common-case setup collapses from this:

let origin = Origin::random().produce();
let client = Client::new().with_origin(origin.clone());
let session = client.connect(transport).await?;
// ... use origin to publish / consume ...

to this:

let session = Client::new().connect(transport).await?;
let pub_origin = session.publisher().unwrap();
let sub_origin = session.consumer().unwrap();
// ... use them ...

with_publish / with_consume / with_origin stay for callers that need a custom topology (relay, gst, cli, ffi); when any of those is set, session.publisher() and session.consumer() return None and the caller drives things through the origin it already holds.

Targeting dev per CLAUDE.md (breaking change to public APIs in rs/moq-net, rs/moq-ffi, and the language wrappers).

Rust layer (rs/moq-net)

Session now carries the auto-created origin sides directly:

pub struct Session {
    /* existing fields */
    publisher: Option<OriginProducer>,
    consumer: Option<OriginConsumer>,
}

impl Session {
    pub fn publisher(&self) -> Option<&OriginProducer> { ... }
    pub fn consumer(&self) -> Option<&OriginConsumer> { ... }
}

Client::connect and Server::accept both auto-create an Origin when neither publish nor consume was set, then surface it through these accessors. One type instead of a separate ClientSession wrapper, symmetric across the client and server paths.

Downstream call sites in moq-native, moq-relay, moq-cli, moq-gst, libmoq, and moq-ffi thread session.publisher() / session.consumer() where they need the auto-created sides; everything else just continues to use session as before.

One internal moq-net test was depending on "no origin → trivial close with code 0 (Cancel)" to validate SETUP framing. With auto-origin attached the Lite01 session now closes with RequiredExtension (code 1) since Lite01 can't satisfy the new origin. Loosened the assertion to "any non-Version error" since the test's purpose is the framing path, not the close code.

FFI layer (rs/moq-ffi)

Mirror at the FFI level: MoqSession gains publisher() and consumer() accessor methods (uniffi Objects, return Option<Arc<MoqOriginProducer>> / Option<Arc<MoqOriginConsumer>>). MoqClient::connect continues to return Arc<MoqSession>.

Renamed MoqOriginProducer::publish → add_broadcast so session.publisher().unwrap().add_broadcast(path, broadcast) reads cleanly instead of stuttering …publisher().publish(...).

New test server_client_roundtrip_auto_origin validates the no-config path end-to-end (client doesn't call set_publish/set_consume; session.publisher() and session.consumer() are Some and drive a server → client broadcast).

Per-language wrapper sync (CLAUDE.md cross-package table)

• py/moq-rs/moq/client.py: no extraction step needed, just stores the returned MoqSession directly.
• py/moq-rs/moq/origin.py: Python OriginProducer.publish wrapper kept (no Python API churn); now calls add_broadcast on the FFI underneath.
• swift/README.md, doc/lib/{swift,kt}/{index,moq}.md: show the new shape — no MoqOriginProducer() construction in the happy path; session.publisher() / session.consumer() return the auto-created sides.
• go/: pure binding shim, no Go-side wrapper to update; consumers get the regenerated MoqClient.Connect() returning *MoqSession with the new accessors.
• Swift Moq.swift / Kotlin Moq.kt: already deleted in the prior PR swift+kt: re-export FFI, session.shutdown(); explicit Origin wiring #1526, so no entry-point shims to update here.

Why fold into Session instead of keeping ClientSession

The earlier version of this PR added a separate ClientSession { session, publisher, consumer } wrapper. Two reasons to fold:

1. Symmetry: the same auto-origin behavior applies to servers too. Keeping ClientSession would force either a parallel ServerSession (more types) or asymmetric APIs. Folding the fields onto Session lets Server::accept use the same path with zero new naming.
2. Naming: there's no obvious neutral name for the wrapper (Endpoint and Connection are overloaded), and Session already represents "the connected thing." Adding optional fields fits cleanly without inventing a new concept.

Test plan

• cargo test --workspace --tests passes (all suites, including the new server_client_roundtrip_auto_origin).
• cargo fmt --check clean; cargo clippy --workspace --all-targets clean.
• just swift check passes (2/2 smoke tests).
• Kotlin: gradle toolchain not available locally, so just kt check only ran the binding regeneration step. CI exercises the full Kotlin build.
• Python: smoke tests not run locally (no uv environment set up in this worktree). CI runs just py check.

(Written by Claude)