- Removes blockUpper2GB hack; we now start early in the process
and reserve all (available) space in the lower 2GB address
space, leaving the upper 2GB untouched for host code
- All virtual memory operations flow through wibo::heap for
bookkeeping
- All guest code uses a guest mimalloc area + thread-local heaps
reserved in the guest address space
`TlsGetValue` disambiguates 0 and an error by relying on `GetLastError`. Depending on the program state, `GetLastError` could be non-0, even though `TlsGetValue` succeeded. Resolve this by always setting `wibo::lastError`. This matches the behavior described by the documentation.
Additionally, when reading resources, later versions of mwcc and mwld call `GetModuleHandleA` with the program path, and then call `LoadStringA` on that handle. Support this behavior by _actually_ loading the PE at the path passed in to `GetModuleHandleA`, instead of assuming it's the current program.
(This is especially useful because sjiswrap relies on overriding `GetModuleFileNameA`, so the wrapped program reads its own resources, rather than sjiswrap's.)
Other small changes:
- Add ms-win-crt `exit` & run atexit funcs
- Implements vcruntime `memmove`
- Implements kernel32 `GetModuleFileNameA`
