fix(json): register provider and fix JSON/YAML parity gaps

[sspaink]

Follow-up to #192, #193 and #194, and the last of the five unregistered providers. The BuiltinProvider entry for opa-builtins-json was commented out in its META-INF/services file, so none of the twelve json/yaml builtins were reachable for consumers.

Registering it exposed 16 failing compliance cases:

• json.marshal_with_options ignored pretty and prefix and used Jackson's layout. Go composes prefix + MarshalIndent(v, prefix, indent): the prefix leads the document and repeats on every line, the separator is ": ", an empty container stays on one line, and an explicit "pretty": false disables indent and prefix. An unknown option key is a type error, which was not reported at all.
• json.match_schema and json.verify_schema accepted a schema with an unknown "type". The validator silently never matches it, whereas Go rejects the schema, so the schema is now checked up front. Their error objects also used networknt's field names and wording rather than the desc/error/field/type shape Go returns.
• json.patch could not address Rego sets. Sets serialize as arrays, so their locations are recorded before serializing: a member is matched by value rather than index, an absent member appends, and the result is rebuilt as a set. Since the path segment is the member, an add whose value differs from the segment is undefined.
• json.is_valid raised a type error for a non-string operand where Go returns false.
• yaml.unmarshal now reports Go's terse yaml: line N: <problem> instead of SnakeYAML's multi-line snippet. The residual wording difference (expected ',' or ']' vs yaml.v2's did not find expected ',' or ']') is recorded in ComplianceTest's alternate-message map, as mapping every SnakeYAML diagnostic onto yaml.v2's phrasing is not tractable.

This also raises the evaluator's test heap. The jsonpatch/json_patch_tests fixture runs the whole upstream JSON Patch spec suite inside a single policy, and once json.patch resolves it needs more than Gradle's default 512m (1g is enough; 2g leaves headroom). That is a test-only limit, but it does say json.patch is memory-hungry on large inputs and is worth profiling separately.

Adds JsonPatchSetTest for the set behaviour, three of whose four cases fail against the unfixed code.

Removes the twelve json/yaml entries from known-missing-builtins.txt (64 → 52). With all five providers registered, every remaining entry is a builtin that is genuinely unimplemented rather than merely unreachable.

Note: #193 and #194 are still open and also edit known-missing-builtins.txt, so whichever of the three merges later will need a rebase.