[REA-5060] Add the managed distributed video model adapter - #176

…#195)

https://linear.app/reactor-team/issue/REA-6343 · Design: [DistributedWorkers](https://app.notion.com/p/reactor-technologies/DistributedWorkers-3decacd6a3968030b598d95d6df32ba0)

First of three. Standalone: nothing else in the runtime imports `reactor_runtime.distributed`, and this stack does not depend on the `ReactorApp` stack (#187 to #192). Lifted from #134, #175, #176, #177, which it closes.

## Why

The runner in #196 moves a model's input and result between processes. Both are the author's own objects and hold large arrays. Pickling them through a queue copies every byte; a fixed frame buffer forces a shape on the author. Neither is acceptable.

## What changed

`pack()` pickles with protocol 5. Each flat buffer (a contiguous numpy array) goes into a shared-memory block; the pickle stays small and rides the queue as a header that names the block and the spans. `unpack()` attaches to whatever block the header names and copies the spans out.

```python
slot = SharedSlot()                  # writer side, one per direction
header = pack(result, slot)          # ~1 KB on the queue; the arrays are in the block

reader = SlotReader()                # reader side
result = unpack(header, reader)      # the same object, arrays copied out
```

The block grows to the writer's high-water mark. Since the header names the block, growth needs no protocol. A torch tensor or a sliced array does not expose a flat buffer and pickles inline: slower, still correct, reported once per process. `SharedSlotAllocationFailed` names `--shm-size` and the pod's `/dev/shm` when a block cannot be created.

Also here: the wire (`Generate`, `Reset`, `Shutdown`, `Loaded`, `Answer`) and the runner's errors (`WorkerCrashed`, `WorkerTimeout`, `RankDesync`).

## Verification

`tests/unit/distributed/test_distributed_ipc.py`: a struct with a dict, str, enum, list, bytes, nested dataclass, `None`, and two arrays of different dtypes crosses intact with both arrays out of band; growth mid-call keeps what was already written and unlinks the old block; a non-contiguous array pickles inline and is reported once; the allocation error names the fix. `mise run lint`, `typecheck`, `test` green.