feat: make the OpenTelemetry sampler configurable, and compress the ADRs

[lesnik512]

Closes #184.

opentelemetry_sampler on OpenTelemetryConfig, handed to TracerProvider(sampler=...). Three
follow-on commits reshape docs/adr/, which is where this PR's own rationale lands.

Sampler, not a sample-rate float

The issue leaned towards opentelemetry_traces_sample_rate: float | None, to keep an
opentelemetry.sdk type out of a config whose OTel extra is optional. That argument does not hold
here: opentelemetry_instrumentors already carries BaseInstrumentor, and sentry_integrations /
sentry_before_send carry Sentry types, all as quoted annotations over TYPE_CHECKING imports that
never run.

What decided it is that a float closes only part of the gap. This instrument owns provider
construction and has no opentelemetry_additional_params escape hatch, so after a float shipped,
ALWAYS_OFF, a bare TraceIdRatioBased without the ParentBased wrapper, and vendor samplers
would all stay unreachable, and a second field would follow. A float can still be layered on top of
this one later; the reverse ordering could not. ADR-0008 records it.

Two premises in the issue that did not survive checking

TracerProvider(sampler=None) is identical to omitting the argument. The SDK does
if not sampler: sampler = sampling._get_from_env_or_default() — at the declared 1.28 floor and at
1.44 alike. No conditional, no sentinel default; the field is passed unconditionally.

Sampling was already configurable, through OTEL_TRACES_SAMPLER / OTEL_TRACES_SAMPLER_ARG.
Precisely because the old code omitted sampler, the SDK read the environment. The gap was the
absence of a Python surface, not the absence of any surface — which makes the env path something the
change has to preserve rather than a non-feature. Two tests pin it: unset still honours the env
vars, and an explicit sampler wins over them.

Both of those also mean the default is unchanged, as the issue asked.

Not covered: FastStream (#223)

The issue expects this to reach every bootstrapper "since they share OpenTelemetryInstrument".
FastAPI, Litestar and Free do. FastStream does not, for a pre-existing reason:
FastStreamOpenTelemetryInstrument.bootstrap() overrides the base without calling
super().bootstrap() and never constructs a provider — it reads get_tracer_provider() and hands
that to the middleware. opentelemetry_endpoint, the resource attributes and
opentelemetry_log_traces are equally inert there today. Filed as #223 with the two options, since
fixing it means lite-bootstrap starting to build and globally install a provider for services that
currently supply their own.

The ADR commits

docs/adr/ had grown a **Decision:** / ## Context / ## Rejected alternatives / **Revisit trigger:**
scaffolding, 3,421 words across ten files. They now follow the /domain-modeling format — a title
and a short body — at 1,423 words across eight, with revisit conditions folded into the prose.

Two merges, no drops. The exclusion-policy ADR joins the per-instrument-axis one: both record a
rejected re-organization refused for the same reason, that it moves complexity rather than
concentrating it. The OTLP-HTTP extra joins the orjson one: both are the same decision applied twice,
a native dependency with no free-threaded wheels becoming an opt-in extra rather than a
combinatorial matrix of variants. The set is then renumbered 0001-0008, so the sampler ADR above is
0008 rather than 0010.

Renumbering breaks citations, and the invariant that was supposed to catch that only read full
docs/adr/NNNN-slug.md paths in Python — of which the repo has exactly one. The other eight
citations are bare ADR-NNNN numbers in comments and prose, plus a path in pyproject.toml; none
were checked, and scripts/floor_smoke.py was left pointing at the wrong ADR until a manual grep
found it. The last commit closes that: the scanner reads .py, .md and .toml and resolves both
forms. Verified by breaking each form in turn.

Tests

Four against the instrument (sampler reaches the provider; unset keeps DEFAULT_ON; unset honours
the env vars; explicit sampler beats them) and one through FreeBootstrapper, so the pass-through is
pinned at the bootstrapper level too. Two more cover the widened citation scanner. 290 pass.