HFIR adapter: exponential lowering on shared pure subexpressions; add depth/expanded-size cap (blocks live runs) #169

From Motoko's #167 review: lowering doubles per level of shared pure subexpression (depth 20 = 2.2s/3.1GB; depth 22, a 3.7KB proposal, runs out of memory). A 4095-node chain allocates about 4.3GB. Memoize/share lowering and add a depth or expanded-size cap that fails closed. Blocks live-model runs.