Move kmath-memory into a separate project

[altavir]

Create a separate project for low-level multiplatform memory management. Let's call it memory.kt. Implement several memory backends:

• ByteBuffer for JVM
• WebGL buffers for JS
• ByteArray for Native
• JEP 393 foreign memory support for JVM (#83). Additional module with JDK 17 target.

The following features should be present:

• Random access primitive read
• Random access primitive write
• Separate lifecycle handlers for readers and writers (control that only one writer could exist at a given time).
• [Optional] Freeze access on borrow via coroutines like suspend fun Memory.borrow{ block: Memory.()->Unit} to pass it to other code or external process
• [Optional] Kotlinx-serialization infrastructure to perform operations directly on memory (could be useful for ktor-io).

[NorbertSandor]

As I see this is already implemented, is it? Thanks.

[altavir]

No, it is not. @CommanderTvis started working on this, but It is not finished. I have some ideas, but it requires a lot of research to understand how it should work and what benefits it has. For example I have an idea about using a Buffer/Matrix/Tensor algebra based on manual memory layout with pooling/nogc implementation of the Memory. But sadly I have no time for it now. All ideas and contributions are welcome.

[NorbertSandor]

@altavir Thanks, but the more I dive into the requirements of the "waiting for a hero" issues, the more I see that these are at another level :D

[altavir]

Yep, it is a research library. It is requires a lot of thinking. For straight cases I would start with https://github.com/mipt-npm/kmath/issues/405 or https://github.com/mipt-npm/kmath/issues/66