Add OTEL trace capture via koltp for LLM requests

## Context

[koltp](https://github.com/kestra-io/koltp) is a cross-platform process wrapper that captures OTEL traces, metrics (CPU/mem/network/IO), and structured logs from any process.

When a task runner has `monitoring.enabled: true`, koltp wraps the process and forwards OTEL data back to Kestra via `::{"oltp":}::` markers on stdout.

## Goal

For `plugin-ai`, koltp will automatically capture all HTTP requests emitted to LLM APIs (OpenAI, Anthropic, etc.) as OTEL traces — including prompt content and response metadata — without any code changes to the plugin itself, as long as the LLM client library uses standard HTTP.

## Work

- Validate that LLM client libraries used in `plugin-ai` emit OTEL traces or are instrumented via auto-instrumentation when an OTEL collector endpoint is available (set by koltp)
- Add documentation / task property notes explaining what traces are captured when `taskRunner.monitoring.enabled: true`
- If auto-instrumentation is insufficient, add explicit OpenTelemetry SDK calls to capture LLM request/response spans

## Value

- Visibility into every LLM call made by a flow: prompt, model, latency, token usage
- Stored inside Kestra's trace store, queryable over time
- Also forwarded to any external OTEL collector if configured

## Related

- koltp: https://github.com/kestra-io/koltp
- System usage tracking issue: https://github.com/kestra-io/kestra/issues/310
- Task runner monitoring prop (to be added per task runner)
- Implementation challenge: https://app.notion.com/p/kestra-io/Implementation-challenge-37c36907f7b580ea9af9c86569af20b3

Start by reviewing the LLM client libraries used in plugin-ai and the task runner monitoring property notes, then check koltp's OTEL endpoint and marker behavior. Validate whether automatic instrumentation captures the requested spans; if not, determine the scope of explicit SDK instrumentation. Done means traces are captured when taskRunner.monitoring.enabled is true and the behavior is documented.

Written by the indexing model from the issue text.