Create a new backend for onednn with a loader for the compiler

[kurapov-peter]

No description provided.

[dchigarev]

• Add elyzor backend in oneDNN with empty partition::compile/execute methods https://github.com/intel-ai/oneDNN/pull/4
• Implement partition::compile/execute methods using graph compiler's API & load implementation using dlopen Use graph compiler's interface for compile/execute at elyzor backend kurapov-peter/oneDNN#2
• Propagate shapes of input logical tensors at partition::compile to the serialized graph graph: backend: elyzor: fill input graph's tensors with shapes before converting to json kurapov-peter/oneDNN#3
• Check for the GC library version at oneDNN's backend side graph: backend: elyzor: verify gc API version when loading '.so' kurapov-peter/oneDNN#5
• Use oneDNN's functionality to extract machine flags (it used to depend on gc's internals) graph: backend: elyzor: use oneDNN's ISA traits to determine cpu features kurapov-peter/oneDNN#4
• Get rid of macroses redefinition in dnnl_graph_compiler.h (context) graph: backend: elyzor: do not mark GC API as DNNL_API kurapov-peter/oneDNN#6
• Adapt oneDNN's tests for gc 1.0 for the new elyzor backend Add sanity tests for elyzor backend graph: backend: elyzor: add sainity tests kurapov-peter/oneDNN#7

[kurapov-peter]

@dchigarev, should we close this one?

[dchigarev]

@dchigarev, should we close this one?

@kurapov-peter

the only thing that's missing in the new backend is this PR kurapov-peter/oneDNN#4 , lets merge it and then we can close this issue

[dchigarev]

All PRs from the list were merged, so closing