Implement linking

[vosen]

Expected outcome:
Compiler exposes APIs that support cuLinkCreate, cuLinkAddData, cuLinkComplete. Enough support for linking PTX. No JIT options support is necessary.

Comments:

• Explicit use of the host-side cuLink* is niche, the same functionality is more commonly used to compile certain fatbins.
• While it's tempting to use LLVM built in linking support, it has subtly different semantics from PTX linking, so the linking has to be handled entirely by ZLUDA
• Implementation should most likely be a pass either right after ptx has been normalized from text operands to integer identifiers, as the first pass or during the normalization. Old ZLUDA did this during the normalization pass, but it has proven cumbersome