Add configurable batch annotation to reduce LLM round-trips during the
semantic annotation pre-pass.

Previously: 36 operations × ~15s per call = ~9 minutes sequential.
With --annotation-batch=10: 4 batches × ~20s per call ≈ ~1–2 minutes.

Design:
- --annotation-batch N (default 0 = sequential, one call per op)
- Each batch sends N operations in one prompt; LLM returns a JSON array
  keyed by operation_id (not array index) for reliable matching
- Partial failure: if a batch returns fewer entries than requested, ops
  without a matching operation_id get no annotation — generation continues
- Batch JSON error: ops in that batch get no annotation, warn logged
- Engine.SetAnnotationBatch(n) for programmatic control

Reliability:
- operation_id-keyed responses survive LLM reordering or omissions
- parseBatchAnnotations silently drops entries missing operation_id
- Invalid JSON returns nil map → all ops in batch get no annotation
- No cascading retry per op (annotation is best-effort by design)

Tests:
- TestEngine_BatchAnnotation_EmitsOneEventPerOp: TUI progress unaffected
- TestEngine_BatchAnnotation_AnnotatesOpsCorrectly: key-based matching
- TestEngine_BatchAnnotation_BatchFailureIsGraceful: invalid JSON safe
- TestEngine_BatchAnnotation_SplitsIntoBatches: 5 ops / batch=3 → 2 calls
- TestParseBatchAnnotations_*: unit tests for the parser
- AT-252: --annotation-batch flag registered in gen --help