Expose the inference pipeline

[bsbodden]

What changed

• add an owning InferencePipeline that exposes tokenizer, metadata, active context state, structured prefill, forward logits, reset, checkpoint, rewind, and normal generation
• expose runtime context capacity independently from the model-declared maximum
• preserve the generation prefix-cache invariant when callers use low-level context operations
• document the advanced inference path and prepare release 0.2.6

Why

The low-level backend primitives existed, but users had no coordinated, discoverable API for the complete inference pipeline. Applications that need llama.cpp-style control should not have to choose between the high-level generation loop and unsafe direct backend access.

Impact

Existing TextGenerationModel and backend APIs remain source-compatible. Applications can opt into InferencePipeline; the pipeline owns and closes its backend.

Validation

• ./gradlew check verifyReleaseMetadata --no-daemon --stacktrace
• ./gradlew verifyPublishingConfigured verifyReleaseDependencies aggregateJavadoc --no-daemon --stacktrace
• ./gradlew :docs:build --no-daemon --stacktrace from a normal clone
• focused Models API, runtime, Java backend, and native backend tests