Enable constant BlockShape requests for NN classification

[FynnBe]

Tiktorch networks expect valid shapes. We cannot meaningfully extend unsuitable requests on the Tiktorch side, as the raw data resides elsewhere (atm with ilastik).

In the current version of the nn workflow these ops are responsable for relevant block shapes:

• OpCompressedUserLabelArray (training samples)
• BigRequestStreamer (headless prediction)
• OpBlockedArrayCache (predictions for GUI)

This should be extended to a list of valid block shapes for better performance, but these are only the edge cases.

• Extend to lists or open new issue for this

[FynnBe]

The caches/BigRequestStreamer should operate with the max block shape. Smaller prediction requests are handled by the classifier: 31c7a90