HermitCrab analysis: finish each stratum as it completes instead of keeping every state until synthesis (per-word memory) #639

HermitCrab analysis keeps every word of every stratum alive until the whole analysis ends, so per-word memory grows linearly with the number of distinct states reached. AnalysisLanguageRule.Apply adds each stratum's outputs to one results set, Morpher.ParseWord copies that into a queue, and only then does synthesis start. The only bound is MaxAlternatives, which counts alternatives, not bytes.

Recorded for later, in the same family as #593 and #595 (which shrink each word; this one is about how many are live at once). C# numbers for this mechanism do not exist yet; the numbers below are PanGloss ones. The C# structure is the same: breadth-first stratum search, a deduplicating output set, equivalents folded into a canonical.

Idea (the PanGloss approach, owner-approved there): finish each stratum's results as the stratum completes (hand its canonicals to synthesis, drop them once the next stratum has consumed them), keep a compact exact visited key per state instead of the whole word, and keep the full search tree only when tracing (Try a Word).

Trade-off: uncapped words give identical analyses. A capped word's partial result changes in PanGloss because work is spent in a different order. In C# the cap throws MaxAlternativesExceededException rather than returning partials, so the outcome is expected to stay "throws", but where it throws and the enumeration order of results would change; that needs checking. The compact key must be exact: a hash-only visited set can lose a parse on collision and is not proposed.

Reference implementation: sillsdev/PanGloss#249 and #250 cover the sharing and alternative-storage parts; the per-stratum finishing is the planned next step there.

Next: recorded for later, not scheduled. Measure C# peak working set on a heavy word (for example Mbugwe mindeshi) after #593 and #595, to see how much remains.