fix: spans built from non-SpanContext parents crash or corrupt trace continuity

[reevejd]

fixes #513

Problem

non_recording_span wraps any object and exposes it as the span's #context, which the
next span start reads back as its parent. Three paths hand it something other than a
SpanContext, and should_sample? then crashes the request with
NoMethodError: undefined method 'tracestate':

• start_span's agent-not-ready path passes the raw OpenTelemetry::Context (the
  agent-warmup crash from the issue)
• the gRPC server instrumentation passes {} for any request without trace headers
• ActionCable stores {} for connections processed outside a trace

non_recording_span(nil) also fabricates a random but valid
SpanContext, so a warmup request with no parent at all records later spans under a trace
id that was never reported.

Fix

I considered just adding a guard like my monkey-patch in the issue, which is enough to
stop the 500s, but it only changes the symptom: malformed parents keep flowing and requests
silently produce broken traces (dropped continuity, fabricated trace ids) instead of
crashing. So this PR fixes the producers and keeps the guard as another layer of defense:

• tracer.rb: the not-ready path extracts the current span's SpanContext from
  with_parent (defaulting to the current context, like the recording path), yielding
  SpanContext::INVALID when there is no parent
• samplers.rb: reads tracestate only when the parent supports it, falling back to
  Tracestate::DEFAULT instead of nil (the SpanContext contract requires a Tracestate)
• grpc.rb: headerless requests use nil instead of {} and start a root span
• action_cable.rb: stores SpanContext::INVALID instead of {}

Tests

Five of the six new tests fail against master, reproducing each defect: the raw-Context
parent, the fabricated trace ids, the sampler NoMethodError, and a headerless gRPC call
that errors the RPC. The sixth pins tracestate propagation (with a tracestate distinct
from the default) so the new fallback can't silently drop real incoming tracestate.

[arjun-rajappa]

Hi @reevejd
Thanks for tracking this down — identifying the root cause and the fix for something that's been reported by so many people is great work, and this looks very promising!

A couple of minor suggestions before we merge:

• The new tests in test/trace/tracer_start_span_test.rb could be moved directly into the main tracer_test.rb file — one less file to maintain.
• Similarly, the sampler tests would fit better under test/samplers/sampler_test.rb

Everything else looks good. I want to spend a bit more time verifying the fix to make sure we're not missing any other scenarios or code paths where this could surface — I'll post an update once I'm done.

FYI - I closed and reopened the PR to trigger the test pipeline.

[reevejd]

Thanks @arjun-rajappa! I've made the changes you suggested.