[Feature] First-class agent targets, not just model endpoints

[xinyu-dev]

Problem

A routing target is modeled as an LLM endpoint. You can point one at an agent, but
only by wrapping the agent in an OpenAI-compatible adapter yourself — and the
defaults still assume a model:

• DEFAULT_DETERMINISTIC_TIER_TIMEOUT_S is 600s, far short of an agent that plans
  and executes code
• nothing streams while an agent works, so client idle timeouts fire on a healthy run
• there is no convention for what an agent target returns (usage, model id)

Why it matters

Routing between models and agents is the interesting case: a cheap model for routine
work, an agent when the request needs tools or code execution. We ran a three-target
route (Claude Opus, a literature-search agent, and BioMni) and every agent-specific
concern had to be solved outside Switchyard — an adapter, a bridge, a credential
proxy, and six hand-tuned timeouts.

Proposed behavior

A target kind that says "this is an agent, not a model":

targets:
  biomni:
    kind: agent
    model: gym/biomni-agent
    base_url: http://127.0.0.1:9300/v1

With it:

• agent-appropriate timeout defaults, instead of a model's 600s
• no assumption that first-byte latency is small
• a documented response contract so adapters do not each invent one

Notes

Related to #240. That issue lets a classifier pick among arbitrary named targets;
this one asks that those targets can be agents rather than models. They are
complementary — #240 is the routing decision, this is what sits at the end of it.

[xinyu-dev]

Related: #240 (classifier selecting among arbitrary named targets), #244 (agent targets), #245 (in-flight target visibility), #246 (classifier truncation), #247 (NeMo Gym agents).