Outside of span streaming, an outgoing fetch (`http.client`) span with no local
parent is no longer recorded as a standalone transaction — the downstream
sampling decision is left to the server. This is enforced via `onlyIfParent`,
which still creates a non-recording span so trace propagation headers are
injected.

This rule already lives in `SentrySampler`, but that only runs when an
OpenTelemetry SDK tracer provider is set up. Enforcing it in the instrumentation
makes it hold for the `SentryTracerProvider` and for SDKs that don't use an
OpenTelemetry tracer provider at all. The sampler rule is kept for OpenTelemetry
SDK / custom OpenTelemetry setups.

…Event hook

`.withResponse()` awaits the instrumented promise (which ends the gen_ai
span) before returning, but `.asResponse()` routed straight to the raw
`APIPromise.asResponse()` and never waited for it. The span then ended on
the instrumentation's own parse schedule, which can be one microtask after
the enclosing transaction has already been assembled. The SentryTracerProvider
assembles transactions synchronously on root-span end (no debounced span
flush), so the unfinished gen_ai span was dropped from the transaction,
orphaning its child `http.client` span.

Mirror the `.withResponse()` handling: await the instrumented promise
before returning the raw `Response`, so the span ends before the caller
continues, deterministically on both the provider and SDK paths.

Applies to both the OpenAI and Anthropic instrumentations (shared util).

These assert prisma's engine spans (replayed asynchronously by
`@prisma/instrumentation`), which the SentryTracerProvider drops because it
assembles transactions synchronously on root-span end with no SpanExporter
buffer to wait for late children. They pass on the OpenTelemetry SDK
(`BasicTracerProvider`) path. Skip them here until the general
"complete span-tree capture without a SpanExporter" follow-up lands; v7 is
left enabled as it currently captures the engine spans in time.