Instantiate DatasetContextClassifier in DatasetParser and guard null case

[lfoppiano]

Summary

• Fix NPE at DatasetParser.java:1485 on every /service/processDatasetPDF request: this.datasetContextClassifier was permanently null because the getInstance(...) factory accepted the classifier as a parameter but the private constructor it delegated to only took configuration, making this.datasetContextClassifier = datasetContextClassifier a field-to-itself no-op. The same pattern was already fixed for DatasetDisambiguator in Fix NPE when disambiguator is null in DatasetParser #18 — this change applies it to DatasetContextClassifier: eager init via DatasetContextClassifier.getInstance(configuration) in the constructor, try/catch so model-loading failures don't cascade, warn-once helper, and a null-guard at the call site for graceful degradation.
• Also removed the equivalent dead no-op line this.dataTypeClassifier = dataTypeClassifier; (real init is lazy in dataTypeClassify at line 649-650 and untouched).
• Bumped project version to 0.9.0 in build.gradle, resources/config/config.yml, resources/config/config-docker.yml, and Readme.md (docker tags, GROBID_VERSION build-arg, sample JSON response, build instructions). GROBID dependencies are already at 0.9.0 since Remove DataseerML, upgrade to grobid 0.9.0 and Gradle 8.5 #5.

Compatibility with recent fixes

• Fix NPE when disambiguator is null in DatasetParser #18 (DatasetDisambiguator): applies the exact same pattern; no conflict.
• Fix Guice MissingConstructor crash on startup #16 (Guice MissingConstructor): controllers still call DatasetParser.getInstance(serviceConfiguration, null, null, null). The three null args were already ignored by the constructor; the classifier is now sourced from configuration directly.
• Preload models at startup and add model-aware /service/health endpoint #14 (Preload + health): GrobidEngineInitialiser.preloadModels() still marks datasets and context_* correctly. On success the later DatasetContextClassifier.getInstance(configuration) call short-circuits via the singleton guard (no double-load). On context-model failure, the try/catch in the constructor prevents false-blaming datasets, and the dedicated preload step re-attempts construction and marks context_* failed accurately in ModelLoadStatus.
• Warn (don't error) when optional services (entity-fishing, glutton) are unavailable #13 (warn-once): new warnContextClassifierNotAvailableOnce follows the established AtomicBoolean.compareAndSet pattern.

Test plan

• ./gradlew compileJava — passes.
• Happy path: with full GROBID install and context models present, start the service (./gradlew run) and curl -F "input=@<sample>.pdf" http://localhost:8060/service/processDatasetPDF. Expect HTTP 200, no NPE in logs, dataset mentions carrying used/created/shared context labels.
• Degraded path: simulate missing context models by renaming context_used/context_creation/context_shared under grobid-home/models/, restart, and resend the request. Expect HTTP 200 with mentions (no context labels), a single warn-once log line, no NPE, and /service/health reporting the context models as failed.
• Version bump visible: /service/processDatasetPDF JSON response shows "version": "0.9.0".

https://claude.ai/code/session_01MLUqfjKtqyvzt9ALYKpshr