AUS-1004: Add JsonRpcLimits with document and batch bounds

[Astn]

Third of the ten 2.0.0 pull requests sequenced in AUS-1001 (https://linear.app/austins-test/issue/AUS-1001). Ticket: https://linear.app/austins-test/issue/AUS-1004. Stacked on #164; retarget to master once that merges.

• JsonRpcLimits(maxDocumentBytes = 4 MiB, maxBatchCount = 1024), Default, Unlimited; Config.SetLimits(limits) and Config.SetLimits(sessionId, limits) (null inherits). Zero disables a field, negatives throw at construction.
• Byte check at every public Process / ProcessAsync entry before any copy, flattening, checked conversion or transcoding; string overloads measure the UTF-8 byte count and size the buffer by it. Batch check after a full parse and before the first dispatch, counting every top-level element; nothing in an over-long batch runs, notification-only batches included.
• Either limit answers {"jsonrpc":"2.0","error":{"code":-32600,"message":"Invalid Request","data":{"limit":"maxDocumentBytes","maximum":4194304}},"id":null} (or maxBatchCount / 1024) with the configured maximum, identical bytes for every serializer (LimitExceededInfo with its own writer), passed through the parse-error handler when one is set, like the empty-batch answer.
• Kestrel's MaxRequestBytes is unchanged and is met first (HTTP 413 / raw abort); a host that raises it gets the core answer.
• Each entry resolves the session handler and the effective limits once and carries them into the core, so the plain path pays the moved lookup, one volatile read and one long compare; no lock, no allocation. The ProcessAsync memory entry keeps its base exception behaviour (faulted or cancelled task, no new exception region on the hot path).
• LimitsTests (16 cases across the three serializers) and two Kestrel cases; README (raw connections, error table, Configuration/Limits, Security), SECURITY.md, CHANGELOG (Added, Security) and docs/upgrading.md.

Verification: Release build 0 warnings / 0 errors; 1216 tests pass on net8.0 and net10.0; check_request_path_sync.py 23 listed uses, none unlisted (no allowlist change); render.py --check agrees.

Reference machine (8 cores / 16 logical, about 20 percent foreign CPU load from other sessions throughout), base 4724600 vs this branch, interleaved pairs, medians with min..max:

Row	Base	Branch	Note
--sync 3 16, 16 threads RPC/s	34.9 M (28.1..35.5)	34.5 M (30.5..36.1)	level
--sync, allocation per numeric request	0 B	0 B	unchanged (StringMe 32 B both)
--scale 3 16 4.0	pass, 16/1 about 7.0	pass, 16/1 about 7.0	level
--async 3 16 sync/None RPC/s	23.7 M (23.4..24.1)	23.8 M (22.4..24.1)	level
--async Task/None	27.4 M (26.5..27.7)	26.6 M (25.9..27.9)	inside spread
--async ValueTask/None	30.6 M (30.1..31.2)	29.5 M (28.8..29.5)	3.8 percent lower in all three pairs of this set; level in an earlier set; same-path rows level; single-thread micro row level. Flagged for the AUS-985 idle re-measure.
--async yield rows and every B/RPC			level / identical
Micro AsyncDispatchBenchmarks.SyncViaProcessAsync (default job, two pairs)	182.2, 176.3 ns	185.0, 181.9 ns	about 3 ns, within stddev overlap
Micro ValueTaskInline	198.6, 181.6 ns	186.5, 188.9 ns	level
Micro DispatchBenchmarks.Add / AddInt	236.5, 212.6 / 169.4, 164.3 ns	210.1, 238.0 / 162.1, 161.3 ns	level

Raw run files are in the session scratchpad (impl/aus-1004-bench*, impl/aus-1004-micro*).