Optimize common sub-targets

[Andrey Smirnov]

Extract common base steps/dependencies into a common LLB node, e.g. in many cases first step is to copy base dependency, we should better extract that as common LLB node.

What is happening right now is that each sub-step tries to copy concurrently hundreds of megabytes of data which basically means our build is more of IO-bound rather than CPU-bound.