Reset shouldSaveState after use; clamp numThreads to corpus size

[pangwangshu]

Part of #219 — correctness finding #5 and the robustness finding about numThreads > numDocs. Two small, independent fixes.

1. shouldSaveState is never reset, so saveSampleInterval is a dead knob.
collectAlphaStatistics() arms WorkerCallable to collect docLengthCounts/topicDocCounts for the next call(). WorkerRunnable and DMRCallable both clear the flag at the end of their run; WorkerCallable — the class that replaced WorkerRunnable and the one ParallelTopicModel actually uses — never did. Once hyperparameter optimization starts collecting, every subsequent iteration kept accumulating instead of only every saveSampleInterval-th one, silently turning the intended thinned samples into one long, highly autocorrelated run. Now reset at the end of call(), after the armed pass has used it.

2. numThreads > numDocs silently serializes the run.
estimate() computed docsPerThread = data.size() / numThreads, which is 0 whenever numThreads exceeds the corpus size, so every thread but the last got zero documents while the last took the entire corpus. Output stayed correct (empty workers contribute all-zero counts to the merge), but with all parallelism lost and numThreads full copies of typeTopicCounts allocated regardless. estimate() now clamps numThreads to the corpus size, with a log warning, before any of that allocation happens.

Added TestWorkerCallable#shouldSaveStateResetsAfterOneCall (arms the flag, calls call() twice, confirms the histogram total after the second, unarmed call matches the first instead of doubling) and TestParallelTopicModelNumThreadsGuard (trains a 3-document corpus with numThreads=16, confirms it's clamped to 3 and the run still produces valid topic assignments).