Borrow deferred containers instead of copying them

[jerrytron]

json_compiler stored each deferred named child by value (std::tuple<json, std::string>), so a container nested d levels deep was copied d times over: compile_container() recursed into the copy and deferred its children the same way.

The source document outlives the whole compile — it is the caller's json, reached by reference through compile() → compile_container() → handle_container_metadata() — so a pointer is safe and the copy was never serving a purpose.

Measured

On an ESP32-S3, peak heap for a compile fell from 16.4x the source JSON's size to 12.05x. On a memory-constrained target that was most of what made compiling on-device impractical; it matters less on a desktop, but it is pure overhead everywhere.

ctest passes and 77 real stories compile to byte-identical output.

[jerrytron]

Pulled in — the branch now points at your 3d5c7a7.

Verified the reference change is output-neutral: built the compiler at bb08955 and at 3d5c7a7 and compiled 70 real stories with each — byte-identical .cdam in every case. ctest passes.

One note in case it saves you a puzzle later: comparing against our own vendored build showed 68 of those 70 differing, which turned out to be the unstable sort in container_hash_t (the subject of #168), not anything here. The differing bytes are pairs swapping places in the container hash table. Independent confirmation that #168 is worth having.