fix(lfx): bootstrap and flush OTLP telemetry on the lfx run path (#14696)

* fix(lfx): bootstrap and flush OTLP telemetry on the lfx run path

`lfx run` executed the flow but exported zero telemetry even with
OTEL_EXPORTER_OTLP_ENDPOINT set. The OTEL_* environment variables are read
only by bootstrap_application_telemetry(), which lfx serve and the langflow
server call at boot and the run command never did. The flow.execute span the
graph emits (correctly labelled protocol=lfx.run since #14403) therefore landed
on OpenTelemetry's no-op proxy provider and was discarded, so every cron job,
CI step or batch worker wired into an APM through `lfx run` was a blind
surface with no warning.

The run command now installs the providers before the run and flushes them
after it, on every exit path: the process is one-shot, the span sits in the
batch processor when the run ends, and a failed run's status=error span is the
one an operator most needs. The result is echoed before the flush so a stdout
consumer is never held behind the export; the flush runs off the event loop as
in lfx serve. No-op when no endpoint is set or lfx lacks the otel extra.

Tests: a loopback OTLP/HTTP collector drives the real console entry point in a
subprocess with the standard env vars and reads the span off the protobuf that
arrived (success and failure paths), with OTEL_BSP_SCHEDULE_DELAY pushed to ten
minutes so only the exit flush can deliver it; plus an in-process ordering test
(bootstrap -> run under protocol=lfx.run -> shutdown) that needs no otel. Both
fail on the unpatched tip. The existing in-memory probe keeps proving the
protocol binding and now says so.

* test(lfx): tighten the run-command telemetry tests per review

- Skip the OTLP wire test on the OTLP/HTTP trace exporter rather than the
  bare opentelemetry namespace, which opentelemetry-api alone satisfies:
  the exporter import brings the proto module the collector decodes with
  and the SDK the subprocess needs, or fails and skips.
- Record typer.echo as a lifecycle event so the tests assert the actual
  sequence bootstrap -> run -> echo -> shutdown (success and RunError
  paths) instead of inferring echo-before-flush from stdout after the fact.

* chore: trigger CI