fix: harden 0.2 non-web consumer readiness

[ChoMinGi]

A standalone non-web Spring Boot app exposed a startup failure before 0.2.0: Spring was introspecting servlet and WebFlux filter return types even when those dependencies were absent.

This keeps core, servlet, and reactive auto-configuration separate so optional web stacks are guarded at the configuration-class boundary. The 0.1.x factory descriptors remain available through a deprecated compatibility facade. The global langfuse.enabled=false switch now also disables model adapters and operational signals when an application supplies its own LangfuseOtel bean.

Two reactor-external Boot 3.5 consumers now cover Spring AI and LangChain4j independently. Each runs both application-owned and standalone OpenTelemetry paths. The standalone path starts a non-web context, invokes a local ChatModel, flushes to a loopback OTLP receiver, decodes the protobuf request, and checks authentication, ingestion headers, service resource, and model attributes. CI runs Spring AI 1.0.9/1.1.8 and LangChain4j 1.0.0/1.18.0. Provider API calls and live Langfuse ingestion are intentionally outside these credential-free smokes.

The public docs now distinguish the published 0.1.1 artifacts from the 0.2.0 candidate. Release preflight rejects candidate wording, incorrect README artifact versions, an undated version heading, or a missing Unreleased section before Central credentials are exposed.

Verified with the reactor clean verify, the quality profile, both example builds, the core prompt consumer, all four framework consumer combinations, and actionlint. The current PR CI run is green.