SameDiff: Implement new graph format (FlatBuffers storage problems to be resolved)

[AlexDBlack]

We use FlatBuffers for storing SameDiff graphs. Here's some issues we need to solve with it (and our use there-of) going forward:

• 2GB limit (we hit this in practice with large models especially with updater state)
  • Related: 32-bit indexing in Java, all ByteBuffer underneath
• Zero copy deserialization of arrays etc not possible
• Named args for ops
  • Currently we store iargs/bargs/targs - which is bad for UX
  • We also store "java reflection" format which is brittle and language specific
• No (official) Swift support

I would also like to clean up the format/fields, it has been built incrementally, and a lot has changed since it was created.

Alternatively, we look at ProtoBuf as the replacement here. But ProtoBuf may be less suitable for mobile than FlatBuffers.

Also: it turns out that FlatBuffers does support String/JSON format, which may be useful for our ops definitions, if we decided to go that way.

---

Edit 2020/05/05: As discussed internally.

Essentially, we want to implement a new format for SameDiff models that avoids some of the problems with the current format, and also allows for things like:

• Memory mapped access to array values
• Model compression (zip for example)
• Improved ops arg serialization (with improved backward compatibility)

[raver119]

Looks like Swift support is going to be added soon: https://github.com/google/flatbuffers/pull/5603

[AlexDBlack]

New design has been created and merged here: https://github.com/KonduitAI/deeplearning4j/pull/467 Now it's simply a case of implementing that design.