Make it easy to extract typed rerun components for .rrd recordings

[polynuma]

Is your feature request related to a problem? Please describe.
reconstructing a typed rr.Image / rr.Pinhole / rr.Point3d from a .rrd file requires too many lines of the code. That is, after querying an .rrd file (load_recording -> view -> select -> read_all), there's not a simple way to deserialize the columnar format into an iterable set of rows of the original type. Here's an example of recreating an rr.Image - the least complicated of all

    rrd = rr.dataframe.load_recording(rrd_path)
    timeline="frame"
    path="/camera/0/color/image"
    view = rrd.view(index=timeline, contents=path)
    table = view.select().read_all()
    frame_col = table[timeline]
    buf_col_name = f"{path}:Image:buffer"
    buf_col   = table[buf_col_name]
    fmt_col_name = f"{path}:Image:format"
    fmt_col   = table[fmt_col_name]
    for row_idx in range(table.num_rows):
        # ---- timestamp or sequence number ----
        ts_scalar = frame_col[row_idx]
        frame_ts = ts_scalar.value
        # ---- zero-copy bytes ----
        inner = buf_col[row_idx][0]
        pixels = inner.values.to_numpy(zero_copy_only=False)
        # ---- width/height ----
        fmt = fmt_col[row_idx][0]
        w = int(fmt["width"])
        h = int(fmt["height"])
        # ---- reshape -> BGRA8 ----
        img_np = pixels.reshape(h, w, 4)
        # yield
        yield frame_ts, rr.Image(img_np)

As you can see, we must query each pyarrow column with hand-crafted deserializer instead of relying on the type automatically having decorators/macros under the hood to handle. Although this is a consequence of an OLAP / columnar storage format, that does not mean users must suffer !

Describe the solution you'd like
Akin to a typical ECS, we should be able to iter the table given a type

def iter_color_frames(rrd_path: str):
  rrd = rr.dataframe.load_recording(rrd_path)
  view = rrd.view(index=timeline, contents="...path")
  for row in view.select(rr.Image):
     ....

Describe alternatives you've considered
there's lots of ways to slice this problem ... but it can be measured by # lines of code for devs

Additional context
Mirror game engines with ECS semantics - like Bevy