Generation-time quality controls: token masks, mid-word continuation, KV snapshot

[FuJacob]

Summary

Adds generation-time quality controls so the app can stop emitting unwanted tokens at the source. All behavior is opt-in or backward compatible; existing single-sequence and multi-sequence paths are unchanged.

• buildTokenMasks runs once per model load and classifies the whole vocabulary. Control, unknown, and unused tokens get a -inf logit bias (via llama_sampler_init_logit_bias in buildSampler) so they can never be sampled as visible text. EOG tokens are deliberately left sampleable so the existing stop check keeps working.
• New single_line flag on SamplingConfig additionally masks line-break tokens.
• setForceWordContinuation(seq, enabled) constrains the first (seed) token of the next generation to continue the current word (masks whitespace-leading tokens for that one token).
• snapshotSize / snapshotSequence / restoreSequence wrap llama single-sequence KV state copy, serialized with the decoder thread (an enabler for future multi-candidate decoding; nothing calls them yet).
• SampleResult.logprob: computeLogprob returns the chosen token's log-probability under the raw model distribution (<= 0), on both the batched sample path and the seed token. Lets the app suppress low-confidence completions.

Validation

swift build
# Build complete!

COTABBY_TEST_MODEL_PATH=<gemma-3-1b-it Q4_K_M> swift test
# Executed 20 tests, with 0 failures   (exit 0)

New tests: no-model guards for setForceWordContinuation and snapshotSize; model-gated tests for first-token word continuation, the snapshot/restore round-trip, and a finite logprob. Existing tests still pass, including testInterleavedMultiSequenceSampling (greedy determinism intact, so the always-on nonprintable mask does not perturb normal sampling). Without COTABBY_TEST_MODEL_PATH, model-gated tests skip and the suite is green, matching CI.

Notes

• The nonprintable mask is always on (no flag) because it only removes tokens that should never be visible text, and it excludes EOG so natural stopping still works.
• Fixed a latent end-to-end test bug: testEndToEndWithModel detokenized tokens[0], which for models that prepend BOS is a control token that renders to 0 bytes. It now detokenizes a content token.
• SamplingConfig gained one trailing field (single_line); all in-repo constructors are updated. SampleResult gained logprob (read-only for callers).