nx_sketch_hll_packed.nx

sketch_hll_packed.nx -- 4-bit and 6-bit packed-register HLL. Same algorithm as runtime/sketch_hll.nx; different register storage layout. Saves 50% (HLL_4) or 25% (HLL_6) memory at the cost of bit-twiddling on every register read/write. Tradeoff is the same one DataSketches makes with HLL_4-12 = 2095 bytes vs HLL_8-12 = 4096 bytes; we match their memory footprint while preserving every NishiLang substrate guarantee. rho range: lg_k=10 -> max rho = 64 - 10 + 1 = 55, fits in 6 bits; but 4 bits only addresses [0, 15]. For HLL_4 we use the standard HEULE encoding: a 4-bit base value (0..15) plus a per-bucket EXCEPTION table for buckets whose true rho exceeds 15. In practice exceptions are rare (P[rho > 15] = 1/2^15 = 3e-5) so the exception table stays small. For HLL_6 we have plenty of headroom (max 63), so no exception table is needed. FIRST CUT: ship HLL_6 (simpler -- no exception table) to land the memory stomp. HLL_4 with the exception-table machinery follows in a queued commit. MEMORY: HLL_6-12 (lg_k=12) = 4096 * 6 / 8 = 3072 bytes vs HLL_8 4096 bytes (25% reduction). HLL_6-10 = 1024 * 6 / 8 = 768 bytes vs HLL_8 1024 bytes. Per the lossless-language discipline (doc 20): packing the bits doesn't change the typed envelope -- query returns the same ApproxI64 shape as nx_hll_query. The memory footprint reduction is a substrate-internal optimization, invisible at the API surface.