Optimize arrays only with a session; keep constructors lazy - #10793

The session-less optimize is removed and optimize(session) is the only entry point, so every rewrite sees the session's registered kernels.

Builtins such as cast and fill_null, and slice, filter, take and apply, no longer optimize the node they build. Instead:

• execute_until runs optimize_recursive with the context's session before it starts, and again over the array execution produced, so canonical results do not keep lazy children such as a filtered validity;
• single-step execution optimizes with the session instead of trying only the static reduce rules;
• whenever a rule rewrites a node, the nodes the rule built are optimized too, while subtrees that already existed are left alone, so inverse rules (slice over get_item and get_item over slice) cannot loop.