feat(plugin): add NeMo Relay dynamic integration

[bbednarski9]

What

Adds switchyard-nemo-relay-plugin, an external NeMo Relay native dynamic
plugin that exports only Relay native C ABI v2 and drives Switchyard algorithms
through libsy::Algorithm::run_stream.

The plugin:

• decodes inbound OpenAI Chat, OpenAI Responses, and Anthropic Messages
  requests with switchyard-translation;
• supports weighted, reproducible random routing and the LLM-classifier router;
• resolves every libsy CallLlm step to a semantic target and asks Relay to
  perform the actual provider dispatch;
• feeds buffered responses, preserved provider stream events, and structured
  provider failures back through libsy until ReturnToAgent;
• preserves Relay-owned retry, trusted fallback, cancellation, and
  first-stream-item commitment behavior;
• emits routing requested, decision, retry, error, and fallback marks from real
  libsy steps;
• ships a v2 manifest, configuration schema, bundle builder, fake-provider
  process E2E, and operator documentation.

It also makes run_stream executor-neutral and lazy while retaining its
optional run observer. libsy no longer requires its caller to provide a Tokio
runtime merely to poll an algorithm stream, which is required when the caller
is a host-managed dynamic-plugin callback.

The crate is a source/build unit (cdylib + rlib, publish = false). The
operator-facing artifact is a release bundle containing the shared library,
materialized manifest, schema, license notices, and SHA-256 digest.

Why

NeMo Relay must remain responsible for provider authentication, transport,
retries, fallback, and observability, while Switchyard algorithms may transform
requests, request one or more LLM calls, inspect their results, and select the
final response. A model-selection-only API cannot express classifiers or future
ensemble algorithms.

Relay native C ABI v2 provides the typed, C-safe provider-dispatch contract
needed to implement that lifecycle without linking the full Relay runtime,
using switchyard-llm-client, or running switchyard-server.

This PR is stacked on and depends on #192. PR #192 is what lets same-protocol
provider stream events survive the required Relay -> libsy -> Relay round
trip without losing unknown fields. Feature changes specific to this PR are
introduced by commits 017bea20 and 2efcb3c4; merge commit 9725b24a syncs
the branch with current main and reconciles the observer-capable run_stream
signature without changing #192.

Related: NVIDIA/NeMo-Relay#594

How tested

• cargo test -p switchyard-libsy
• cargo test -p switchyard-translation --test stream_translation
• cargo test -p switchyard-nemo-relay-plugin
• cargo clippy --workspace --all-targets -- -D warnings
• cargo fmt --all --check
• RUSTDOCFLAGS="-D warnings" cargo doc -p switchyard-libsy -p switchyard-nemo-relay-plugin --no-deps
• Ruff and mypy clean for the new bundle/E2E Python helpers
• Manual process smoke: packaged the dylib, validated/added/enabled it with
  the Relay CLI, and exercised buffered plus streaming OpenAI Chat, OpenAI
  Responses, and Anthropic Messages routes
• Manual random-router smoke: all three target protocols, cross-protocol
  routes, unknown-field replay, 12 concurrent runs, retry-once, retry
  exhaustion, and exactly-once fallback
• Manual LLM-classifier smoke: classifier CallLlm, selected-target
  CallLlm, and final response for buffered and streaming requests
• Revalidated after syncing current main: 201 libsy tests, 5 plugin tests,
  focused Clippy, and the cross-repository Relay process E2E pass. The E2E
  reports 12 concurrent random calls across all three protocols, 27 routing
  decisions, 2 LLM-classifier decisions, 2 retry attempts, exactly 1 fallback
  call, and successful buffered plus raw-stream preservation.
• Live Inference Hub smoke: eight buffered and four streaming calls routed
  across nvidia/meta/llama-3.2-1b-instruct and
  nvidia/mistralai/mistral-7b-instruct-v0.3
• Phoenix verification: one complete 14-span OpenInference trace contains
  12 switchyard.routing.requested events, 12
  switchyard.routing.decision events, and both semantic targets

Live trace:
http://10.193.162.160:6006/projects/UHJvamVjdDoxOA==/traces

The repository-wide Rust workspace run reaches an existing macOS
switchyard-py link failure for unresolved Python symbols. All touched Rust
crates and the cross-repository process E2E pass independently.

Checklist

• One class per file; filename = snake_case of the primary class (not
  applicable to the Rust crate; Python helpers follow snake_case).
• New public symbols exported from switchyard/__init__.py.__all__ if
  intended for downstream use (no Python public symbols added).
• Unit tests added for new components / bug fixes.
• README / --help updated if customer-facing surface changed.
• Commits signed off (Signed-off-by: Your Name <email>) per the DCO.

Notes for reviewers

Please review the integration boundary rather than the already-reviewed #192
stream-preservation implementation:

caller wire request
  -> switchyard-translation decode
  -> libsy run_stream
  -> CallLlm
  -> Relay ABI v2 provider dispatch
  -> switchyard-translation decode response/event
  -> CallLlmRequest::respond
  -> ReturnToAgent
  -> switchyard-translation encode to caller

Intentional constraints:

• v2-only: no Relay native C ABI v1 compatibility mode;
• no Relay codecs, private dispatch headers, Relay runtime dependency,
  Switchyard LLM client, or Switchyard service;
• no decision-only/observe-only API or synthesized decision telemetry;
• translation rejects lossy required cross-protocol features rather than
  weakening the request;
• the Relay SDK is pinned to the ABI-v2 feature commit until
  nemo-relay-plugin publishes its first compatible version;
• release automation is deferred until the SDK dependency can move from the
  development commit to that published version.