add methods to read and write serializable objects as datablocks

[axtimwalde]

• introduces DataType.OBJECT that maps to bytes
• introduces block mode 2 because serialized objects do not have a blocksize or
  number of dimensions

[joshmoore]

@axtimwalde: oddly enough I found myself wanting to read a complex type via n5(-zarr) this morning:

colormap-like array-type

  "dtype": [
    [
      "label-value",
      "<i8"
    ],
    [
      "r",
      "<i8"
    ],
    [
      "g",
      "<i8"
    ],
    [
      "b",
      "<i8"
    ],
    [
      "a",
      "<i8"
    ],
    [
      "my-type",
      "<U20"
    ],
    [
      "my-id",
      "<i8"
    ],
    [
      "description",
      "<U200"
    ]
  ],

sample python code

import numpy as np
import zarr

t = [
    ("label-value", int),
    ("r", int),
    ("g", int),
    ("b", int),
    ("a", int),
    ("object-type", "U20"),
    ("object-id", int),
    ("description", "U200")]

data = list()
for x in range(100):
    data.append((1, 100, 100, 100, 100, "Mask", 123456, "some text here"))
    data.append((2, 200, 200, 200, 200, "Mask", 567896, "some more text"))
a = np.array(data, dtype=t)

z = zarr.open("s.zarr")
z.array(name="a", data=a, chunks=(10,))

Is this generally the type of Object you are looking to support here?

[axtimwalde]

No, I think we would have to support this in n5-imglib2 as complex types are just two other types. This is meant for serializing something big into a block, i.e. I need it for an array of sets of point matches.

[axtimwalde]

You can already store arbitrarily many values into a block. It is one of the places where the zarr and n5 specs have never converged. Let me see what we can do for complex types or generally multi-primitive value types in n5-imglib2.

[axtimwalde]

@joshmoore, looking at this again, I am not sure about the best way to support the full dtype language for arbitrary structs in Java as we're not dynamically resolving type information. So far, data of unknown dtype specifications will be opened as byte[] which, in principle, enables interpreting it on the fly with either ImgLib2's ByteArray access or via something similar with ByteBuffer. If you're staying inside the Java language, you can indeed save collections of complex types with this block-serializer but that, in turn, is not very useful when reading from Python.