feat: mirror the agent name onto gen_ai.agent.name on LLM spans

[johndemic]

Closes #126.

Problem

LLM spans carry the agent under the OpenInference key only:

agent.name = build        # opencode's internal agent (build / plan / a subagent name)
agent.type = primary

A consumer reading the OpenTelemetry GenAI conventions therefore sees the
provider, model and token usage but no agent attribution — it cannot tell a
build request from a plan request, or from a named subagent.

Change

Mirror the value onto gen_ai.agent.name, using ATTR_GEN_AI_AGENT_NAME from
@opentelemetry/semantic-conventions (already a dependency). Nothing is
renamed or removed, so OpenInference consumers are unaffected.

agent.type has no GenAI equivalent and is left as-is.

Scope

LLM spans only — startMessageSpan and the completion path in
handleMessageUpdated — matching the issue. Tool, subtask and session spans
also set the OpenInference key and could follow if that is wanted; happy to
extend this PR or send a second one.

A distinction worth preserving

Some backends use a field called agent for the coding assistant itself
(claude-code, gemini-cli, …) rather than an agent within it. This value is
not that — it is opencode's internal agent — so it deliberately is not mapped
onto a bare agent key.

That identity already travels correctly on the resource as
service.name = opencode. Anyone whose backend keys on a bare agent
attribute can set it today with no code change:

"spanAttributes": "agent=opencode"

Tests

Two added to tests/handlers/spans.test.ts; both fail against unpatched
source. They assert the literal attribute string rather than the imported
constant, since that string is the wire format consumers match on — a rename
upstream should fail the suite rather than silently change the payload.

baseline main: 326 pass
this branch:   328 pass, 0 fail

bun run lint, bun run check:jsdoc-coverage, bun run typecheck all pass.

Relationship to #125

Independent — this branches from main and can merge in either order. Both
touch adjacent lines in message.ts, so whichever lands second may need a
trivial rebase.