feat: record time-to-first-token on streaming generations

[PabloPardoGarcia]

Implements GLA2-175.

What

Generation.record_first_token() — call it from a streaming loop when the first content chunk arrives:

gen = glassflow.start_generation("chat", model="gpt-4o")
stream = client.chat.completions.create(..., stream=True)
for chunk in stream:
    gen.record_first_token()   # idempotent — safe to call per chunk
    ...
gen.end()

It adds a gen_ai.first_token span event (constant in semconv.py). Idempotent (only the first call records), no-op on ended/non-recording spans.

Why an event, and why this name

• The backend needs no schema change: the spans table already stores span events, so TTFT is derivable today as event timestamp − span Timestamp.
• No client-side semconv exists for this (OTel standardizes only the server-side gen_ai.server.time_to_first_token metric, emitted by model servers); the event name follows the gen_ai.* style per our convention-native rule. Design precedent: Langfuse's completion_start_time.

Auto-instrumented (OpenInference) traffic — deliberately NOT normalized here

Surveyed all five registry instrumentors:

• openai: already emits its own "First Token Stream Event" on the first streamed chunk (_stream.py).
• anthropic / langchain / litellm (and llama-index): no first-token marker at all — their streaming traces have no TTFT until upstream adds it or we wrap provider streams (the ticket's follow-up).

So TTFT derivation must coalesce both event names. That normalization belongs backend-side (same bucket as GLA2-133's llm.token_count.* gap): collector users and other SDKs bypass our exporter anyway, so an SDK-side rename would duplicate work the ingest layer needs regardless. Reference derivation:

SELECT GenAiRequestModel,
       quantile(0.95)((
         toUnixTimestamp64Nano(Events.Timestamp[arrayFirstIndex(n -> n IN ('gen_ai.first_token', 'First Token Stream Event'), Events.Name)])
         - toUnixTimestamp64Nano(Timestamp)) / 1e6) AS ttft_p95_ms
FROM spans
WHERE arrayExists(n -> n IN ('gen_ai.first_token', 'First Token Stream Event'), Events.Name)
GROUP BY GenAiRequestModel

Testing

TDD: 4 new tests (event lands between span start/end, idempotency, no-op after end(), manual lifecycle) written first and watched fail. Full suite 121 passed; ruff + format + mypy --strict clean.