Allow theta sketches with lg_k = 4 to match Java

[SavicStefan]

Java allows theta and tuple sketches down to lg_k = 4 (16 nominal entries), but C++ rejected anything below lg_k = 5, so the two libraries disagreed on the smallest sketch you can build.

The C++ theta implementation used a single constant, theta_constants::MIN_LG_K = 5, for two distinct roles: the floor on the user-requested nominal size (K) and the floor on the internal hash table (cache) size. Java keeps these separate ThetaUtil.MIN_LG_NOM_LONGS = 4 and MIN_LG_ARR_LONGS = 5, so Java accepts lg_k = 4 (nominal 16) while C++ rejected it in theta_base_builder::set_lg_k.

Because C++ used 5 for both, theta_base_builder::set_lg_k(4) threw, and there was no way to build a theta/tuple update sketch or union at lg_k = 4.

This only affected construction. Reading was already fine: the compact serialization format does not store lg_nom, so C++ has always been able to deserialize Java-produced lg = 4 compact sketches.

Fix, split the one constant into the two that Java has:

• MIN_LG_K -> 4, the nominal floor checked by set_lg_k.
• new MIN_LG_ARR -> 5, the internal hash-table floor used by starting_sub_multiple().

A lg_k = 4 sketch now starts with lg_cur_size (5) > lg_nom_size (4): the hash table keeps its full 32 slots and simply rebuilds down to the nominal 16, a path the existing update/rebuild logic already handles. There is no format or behavior change for lg_k >= 5.

The fix covers tuple sketches as well as theta, with no tuple-specific change. Tuple sketches have no size constants of their own their builder inherits theta_base_builder (sharing the same set_lg_k validation and starting_lg_size()), and their internal hash table is a theta_update_sketch_base (the exact table this PR resizes). So lg_k = 4 now works for tuple update sketches and unions too.

Added new min lg_k tests in both theta_sketch_test.cpp and tuple_sketch_test.cpp. Each test:

• builds at lg_k = 4 and confirms set_lg_k(MIN_LG_K - 1) throws;
• verifies the new MIN_LG_ARR floor -- after many updates the peak retained count is > 16 (2^MIN_LG_K) yet <= 32 (2^MIN_LG_ARR), i.e. the table keeps its 32 slots rather than collapsing to the nominal 16;
• confirms estimation mode, with the true count inside the 2-standard-deviation confidence bounds;
• trims down to exactly the nominal 16 entries;
• round-trips through compact serialize/deserialize.