feat: optional packed J layout to cut JSON memory on small MCUs

[rayozzie]

We are running out of RAM on small MCU hosts, and the JSON object model is where it goes. A parsed document costs 5–9× the size of the text it came from, because every object member is three separate heap allocations — the node, the key, and the value string — and each one pays the allocator's per-chunk header and rounding.

NOTE_C_STORAGE_OPTIMIZATION packs a member's key and short string value into the node's own allocation, so that member costs one allocation instead of three. Members are also reordered to remove alignment padding, taking a node from 48 to 40 bytes on a 32-bit target. valueint and valuenumber are unused on a string node, so their bytes carry the value and then the key; anything that doesn't fit still goes to the heap, and a setter that needs those bytes back evacuates first.

Backward compatible. The historical layout is untouched and remains the default, so nothing shifts unless you opt in. Both layouts render byte-identical JSON from one shared parser and printer. CI builds and tests both, crossed with single precision and low memory.

Results

Measured on a 32-bit target with newlib dlmalloc (4 B header, 8 B align, 16 B minimum chunk).

JSON object	text	objs	allocations	bytes	saved
{"req":"card.version"}	22	2	5 → 2	152 → 96	37%
{"req":"note.add","file":"data.qo","sync":true}	47	4	11 → 4	304 → 200	34%
{"req":"hub.set","product":"com.blues.airnote",…}	93	6	17 → 6	472 → 312	34%
{"req":"note.add","file":"air.qo","body":{6 readings}}	122	10	23 → 10	736 → 536	27%
{"req":"card.location","status":"GPS updated",…}	91	6	17 → 6	472 → 288	39%
{"err":"note: no notes available {note-noexist}"}	49	2	5 → 2	176 → 120	32%
{"req":"env.get","name":"monitor-pump","text":"enabled"}	56	4	13 → 4	328 → 192	41%
{"device":"dev:864475044204278","sn":"pump-A17",…}	75	4	13 → 4	336 → 208	38%
{"a":"b","c":"d","e":"f","g":"h","i":"j","k":"l"}	49	7	25 → 7	584 → 336	42%
TOTAL	604	45	129 → 45	3560 → 2288	36%

Best case is many short string members: 25 allocations become 7. Worst case is numeric bodies, where there is no string value to absorb into the node, but the key still packs.

Over a 50-document corpus the totals are 38% fewer allocations and 26% less heap, dropping expansion from 9.0× to 6.7× the JSON text.

Under NOTE_C_SINGLE_PRECISION the saving is 16%, not 26%: a float JNUMBER already removes the historical layout's tail padding, so both layouts are 40 bytes and the win comes only from packing.

Also fixed

Two pre-existing printer defects, found by comparing output against a pristine build of the base revision:

• Strings containing more than one escaped control character were truncated and emitted invalid JSON.
• Formatted + omitempty printing emitted indentation for members it then elided.

The original idea that I gave to Claude to implement