Add search_json(): query a JSON document given as a string, with optional native accelerator

[SereinCin]

Summary

Adds jmespath.search_json(expression, data) — search a JSON document given as a string.

The common workflow for CLI tools, streams and HTTP handlers is:

import json, jmespath
jmespath.search(expr, json.loads(body))

search_json() does exactly that, and when the optional native accelerator aero-jmespath is installed it runs the whole parse + eval + serialize pipeline in native code, skipping the Python-object round-trip:

jmespath.search_json(expr, body)

This is purely additive and backward compatible: without the accelerator the function falls back to json.loads() + search(), so no behaviour changes and no new dependency is required. Error behaviour is identical in both paths (ValueError for bad JSON, ParseError for a bad expression).

Benchmark

One-shot string-input pipeline (parse + eval + serialize per call), query servers[?status == \active`].instances[0].id on a synthetic server fleet, Windows 11 / Python 3.12. Fully reproducible with the public benchmark repo [aero-jmespath-bench](https://github.com/SereinCin/aero-jmespath-bench) (python run_all.py`):

document	jmespath: json.loads + compiled.search	search_json with aero-jmespath	speedup
0.71 MB	6.47 ms	4.42 ms	1.46x
3.57 MB	42.63 ms	24.71 ms	1.72x
7.19 MB	100.50 ms	51.18 ms	1.96x

The speedup grows with document size. Beyond this one-shot case the accelerator also offers a batch path (aero_jmespath.multi) that parses one document once and evaluates many expressions against it (~1.75x in the benchmark), and its native JSON parser measures ~180 MB/s, faster than orjson (1.36x) and json.loads (1.70x) on the same document.

The accelerator is a native (Aero-compiled) JMESPath kernel that passes all 578 official compliance tests.

Tests

tests/test_search.py gains a TestSearchJson case class covering field access, projections, empty documents, invalid JSON, and matching search() on parsed documents. Full suite: 997 passed.