[AI-383] Add replay-safe OpenTelemetry meter and logger providers - #1710

At worker configuration, warn when the global OpenTelemetry meter or
tracer provider is not replay-safe, pointing users at
ReplaySafeMeterProvider and create_tracer_provider. Add a regression
test proving 1 real execution + 3 replays leaves ADK metric instruments
at their nonzero baseline with ReplaySafeMeterProvider installed, plus a
control asserting the 4x inflation without it.

ReplaySafeMeterProvider unconditionally forwarded get_meter attributes
(added in opentelemetry 1.26) and create_histogram
explicit_bucket_boundaries_advisory (added in 1.30), raising TypeError
into caller code on older APIs within the supported range. Forward them
only when non-None, matching unwrapped-caller behavior.

Validate the global logger provider alongside meter and tracer, and run
the validation from configure_replayer too since Replayer replays are
exactly where unsafe providers re-record telemetry. Warn only on
providers positively identified as replay-unsafe (OTel SDK providers
used directly) so custom wrappers around replay-safe providers no
longer trigger false positives, and compute the warning stacklevel
dynamically so attribution survives wrapping plugins. Add log-event
replay regression tests mirroring the metrics ones.

Logger is an ABC whose __init__ records the instrumentation scope, so the
wrapper now threads name/version/schema_url from get_logger through to the
base class instead of relying on __getattr__ delegation for that state.
Fixes the basedpyright reportMissingSuperCall error that failed lint. Also
documents that opentelemetry._logs is the import path OpenTelemetry itself
sanctions for the logs bridge API while it is pre-GA.

The plugin now warns only when the global meter or tracer provider is an
instance of the public OpenTelemetry SDK provider classes; everything else
(unset proxies, no-ops, unknown wrappers) stays silent, which drops the
private _ProxyMeterProvider/ProxyLoggerProvider imports. The SDK logger
provider is not checked because its class is only importable from the
underscore namespace opentelemetry.sdk._logs while OTel logs are pre-GA.
The opentelemetry.sdk import is guarded for the SDK-less install case, the
warning stacklevel walk uses inspect.currentframe, and the conftest OTel
global resets are documented as the isolation pattern OpenTelemetry's own
test suite uses.

The tracer provider has always forwarded the get_tracer attributes
parameter (added in opentelemetry 1.26) unconditionally, so no release of
this contrib ever worked for tracing below 1.26. Declaring 1.26 keeps the
floor honest at the call level and removes the three conditional forwards
for the attributes parameter; the version-gated forwards for parameters
newer than the floor (context 1.28, histogram advisory 1.30, emit fields
1.38) remain.

The Logs API specification is stable; what keeps the import private is
that opentelemetry-python has not promoted opentelemetry._logs to a
public namespace.

The Span ABC ships add_link as a non-abstract warn-and-no-op default
that __getattr__ cannot intercept, so links added through the wrapper
after span creation were silently dropped.

OTel adds API surface as non-abstract no-op defaults, so a wrapper
method left un-overridden silently swallows telemetry instead of
delegating. Enumerate each wrapped ABC's public and abstract surface by
reflection so new opentelemetry-api surface fails loudly.

The OTel globals are set-once per process and xdist worksteal ignores
xdist_group pinning, so tests share workers with arbitrary siblings.
Assert that global provider installs take effect, park proxy meters on
a no-op provider around the meter-reset fixture, and uninstrument
GoogleADKInstrumentor after test_single_agent_telemetry.