add distributed tracing core (std.trace)

[kacy]

summary

second phase of observability: spans. a span records one unit of work (name,
timing, attributes, status) and its place in a trace. spans nest through a
per-os-thread "current" span, so starting a span inside another parents it
automatically — no context threaded through calls:

span := trace.start("handle_request")
span.set_attr("http.route", "/api")
span.end()

built on pieces already in place: the per-thread current span uses a
threadlocal global (#438); ids come from log.new_trace_id/new_span_id;
time.mono_nanos gives a precise duration on a wall-clock base. tracing is OFF
until set_active(true) (std.obs.init will flip it from the environment), so
start() is a cheap no-op in an uninstrumented program.

• across a spawn: capture current_context() and re-establish it in the task
  with with_context() (the one place propagation is explicit)
• across a process: W3C traceparent format/parse for http/grpc headers
• ended spans collect in a bounded buffer for the OTLP exporter (phase 3) to drain

what was tested

• 3 unit tests: nesting via the current span, no-op when inactive, traceparent
  round-trip
• a cross-spawn program confirming a child task joins the parent's trace and
  parents under the root
• 85 examples

pure-additive stdlib; no compiler change, no bootstrap seed refresh.