Drop the unwind tables the COFF JIT cannot relocate

[tamnd]

Fixes #323.

With #322 merged, mojo run on Windows resolves every symbol it needs and then fails in the linker instead:

JIT session error: In graph , section .pdata: relocation target 0x22cc7961000 (.text) is out of range of Pointer32 fixup at address 0x22cc7961000 (<anonymous block> @ 0x22cc79602b4 + 0x0)
error: Failed to materialize symbols: { (exec, { main }) }

Every x86_64 Windows function carries a .pdata entry describing how to unwind out of it, and each entry is three offsets from the image base. A linker writes one by subtracting the address of __ImageBase from the address of the thing being described, which is what IMAGE_REL_AMD64_ADDR32NB asks for and what JITLink models as a Pointer32NB edge. JIT'd code is not in an image, so COFFLinkGraphBuilder adds __ImageBase as an undefined external, no COFF platform is ever created to define it, and its address stays zero. The lowering pass still runs the subtraction, gets the untouched 64 bit address back, and hands it to a 32 bit fixup. The first function with unwind info trips it, which is all of them.

So this adds an object linking layer plugin that removes .pdata and .xdata from COFF link graphs. It goes in at the front of the pre prune passes, ahead of JITLink's own SEHFrameKeepAlivePass, because that pass gives every .pdata block a keep alive edge to the code it describes and those edges would otherwise be pointing at blocks that are no longer there. .xdata goes too: it is only ever reached through .pdata and it carries the same kind of image relative pointer.

The tables were not being used. Handing unwind info to Windows is the ORC COFF platform's job, we do not create a platform, and the execution engine destructor's platform shutdown branch is unreachable for the same reason, so the entries were being relocated for nobody. What this gives up is walking out of a JIT'd frame in a Windows debugger, which does not work today either. If the COFF platform is ever set up properly the plugin should come back out.

Everything is behind a COFF check, so the ELF and Mach-O paths are untouched.

Part of #31.

[tamnd]

Cross built and checked on a Windows host.

=== version
Mojo 1.1.0.dev0 (deadbeef)
VER_RC=0
=== run
hello from windows
RUN_RC=0

That is mojo run on Windows working end to end for the first time. The .pdata relocation failure is gone and the program runs.

A larger program that exercises List, Dict, string building, range and try gets further and then stops on a different thing, JIT session error: Symbols not found: [ snprintf ]. That is not this change, it is the next one along: snprintf and the rest of the stdio family are inline functions in the UCRT headers and ucrtbase.dll does not export them, so there is nothing on the machine for the JIT to find. Filed separately.