CoreAISequentialEngine: idempotent generation via external session state

[stikves]

Summary
Move per-request generation state off the engine into a caller-owned object, and add a protocol that drives generation over that object. The existing generate() API keeps its current signature. This is the interface a future scheduler needs to run several sessions on one engine.

The new interface

func generate(
  with input: [TokenId],
  sessionState: GenerationSessionState,
  samplingConfiguration: SamplingConfiguration,
  inferenceOptions: InferenceOptions
) async throws -> OutputSequence

augments the older one with explicit sessionState.

Background
The engine kept each conversation's state on the instance: the KV cache, optional persistent (hybrid) states, the processed-token cursor, and the token history used for prefix reuse. The iterator mutates that state as a side effect, so one engine serves one conversation at a time and prefix reuse only applies to the request that ran last. That blocks batched or multi-session serving at the engine boundary.

Changes
GenerationSessionState holds the relocated state (KV cache, additional states, recurrent flag, cursor, history).

It is a class so the iterator's cursor and history writes stay visible to the shim across calls; a struct would copy them and lose prefix reuse.

It is not Sendable: it holds non-Sendable NDArray state and stays in the engine's one isolation domain, so region isolation covers the async generate() call.

IdempotentEngine refines InferenceEngine, the same shape as ConstrainedGenerationCapable: makeSessionState() plus generate(with:sessionState:). It is detected with engine is any IdempotentEngine. Only CoreAISequentialEngine conforms.

The engine now routes every forward-path, reset, and prefix-resolution site through the passed session. The public generate() is a shim over one internal session, which keeps today's prefix reuse, the GenerationTokenBox single-active-generation contract, and reset(to:) semantics as they were.

The prefix-routing decision (full reset on divergence, full reset for hybrid models, rewind on pure extension, and the matching prefix-hit counts) moves into a pure prefixResetPlan() function so it can be unit-tested.

Tests
Builds clean with no new warnings. All new and existing tests pass.

Limitations
Sessions run one at a time. lastPrefixHitCount and the shared GenerationTokenBox stay on the engine, so a concurrent scheduler would need to move or disable those too.

(I would later suggest moving them).