feat: transparent leader hint reconnect on consume

add batch_enqueue() for explicit multi-message RPCs, smart batching via
BatchMode (AUTO/DISABLED/Linger) that routes enqueue() through a background
batcher thread, and delivery batching that unpacks ConsumeResponse.messages
repeated field. update proto to include BatchEnqueue RPC and ConsumeResponse
batched messages field. single-item optimization uses singular Enqueue RPC
to preserve error types. close() drains pending messages before disconnecting.

Update Python SDK for the unified proto API:
- Enqueue RPC now uses repeated EnqueueMessage/EnqueueResult
- BatchEnqueue RPC removed; enqueue_many() replaces batch_enqueue()
- Ack/Nack use repeated AckMessage/NackMessage with per-message results
- ConsumeResponse only has repeated messages field
- BatchMode renamed to AccumulatorMode, BatchEnqueueResult to EnqueueResult
- BatchEnqueueError renamed to EnqueueError
- Linger.batch_size renamed to Linger.max_messages
- Per-message error codes mapped to typed SDK exceptions
  (e.g. ENQUEUE_ERROR_CODE_QUEUE_NOT_FOUND -> QueueNotFoundError)
- All 31 tests pass (16 unit, 15 integration)

feat: 30.2 — unified api surface

- remove unused imports (EnqueueError in client/async_client, Any in
  batcher, threading in test_batcher) — fixes ruff F401
- move MessageNotFoundError/RPCError to top-level imports, eliminating
  inline imports that triggered ruff I001 (unsorted import blocks)
- move ConsumeMessage/EnqueueResult into TYPE_CHECKING block in
  async_client — fixes ruff TC001
- fix EnqueueError docstring to reflect that it is also raised as
  fallback for per-message errors via _map_enqueue_result_error
- strengthen test_flush_single to assert actual request content sent to
  Enqueue, not just call count
- downgrade GRPC_GENERATED_VERSION from 1.78.1 to 1.78.0 in all
  generated grpc stubs (grpcio 1.78.1 was yanked from PyPI)

replace the entire grpc transport layer with a native fibp implementation.
the sdk now communicates directly over tcp using the fila binary protocol,
removing the grpcio and protobuf dependencies entirely.

- add fila/fibp/ module: primitives (Reader/Writer), opcodes, codec
- add fila/conn.py: sync Connection and async AsyncConnection classes
- rewrite client.py and async_client.py to use fibp connections
- rewrite batcher.py to use Connection instead of grpc stubs
- rewrite errors.py with fibp error code mapping (18 error codes)
- add admin methods: create/delete queue, stats, config, redrive
- add auth methods: create/revoke api key, acl management
- add new error types: UnauthorizedError, ForbiddenError, NotLeaderError, etc.
- add ConsumeMessage fields: weight, throttle_keys, enqueued_at, leased_at
- delete fila/v1/ (generated protobuf) and proto/ directories
- remove grpcio/protobuf from dependencies, bump version to 0.3.0
- add test_fibp.py with 34 codec/primitives unit tests
- rewrite test_batcher.py for fibp mock connections
- update integration tests and conftest.py for fibp

the initial codec used generic string maps for admin frames, but the
protocol spec uses typed fields. this aligns all encode/decode functions
with the actual wire format from docs/protocol.md:

- create_queue: [string name][optional on_enqueue][optional on_failure][u64 timeout]
- get_stats_result: typed fields (depth, in_flight, etc.) not a string map
- list_queues_result: [u8 error][u32 nodes][u16 count][per: name, depth, ...]
- set_config: [string key][string value] (not queue + map)
- redrive: [string dlq][u64 count] (not source + dest + u32)
- create_api_key: [string name][u64 expires][bool superadmin]
- set_acl: [key_id][u16 count][per: kind, pattern] (not patterns list)
- all result frames now decode their error_code prefix

also updates types.py with properly structured StatsResult, QueueInfo,
AclEntry, ApiKeyInfo to match the wire format fields.

- add pushback buffer to Connection/AsyncConnection for subscribe() to
  handle servers that send Delivery directly without ConsumeOk
- fix mypy errors: annotate struct.unpack_from return values, fix
  _request_with_leader_retry return type to FrameHeader, move ssl import
  to TYPE_CHECKING block, remove unused make_ssl_context function

the hot-path opcodes were out of order. the server uses sequential
assignment starting from 0x10: Enqueue(0x10), EnqueueResult(0x11),
Consume(0x12), ConsumeOk(0x13), Delivery(0x14), CancelConsume(0x15),
Ack(0x16), AckResult(0x17), Nack(0x18), NackResult(0x19).

- add bounds checks in Reader.read_string() and read_bytes() to detect
  truncated frames instead of silently parsing corrupt data
- add fallback in batcher _flush_many() to fail unresolved futures when
  server returns fewer results than items sent
- remove diagnostic ConnectionError wrapper in _consume_iter that would
  silently swallow decode failures
- tighten subscribe() pushback to only accept Delivery frames, raise
  ConnectionError for unexpected opcodes