Unify handling of compressed init data and use regular tar + gzip to achieve it
There are two issues this commit solves.
First is that architecture-specific code duplicates most of the init binary
handling in each architecture, each with miniscule and confusing variations.
After this commit, the init binary expansion is almost entirely handled by
unified generic code.
Second is that the way we used to generate the incorporated data is somewhat
convoluted. Previously we have a Python script which generates a zip archive
with individual deflate-compressed files and accompanying header and C files
which contain structures describing the archive contents.
The zip file is then extracted and the individual deflate-compressed files are
included in the binary via assembler code.
Since gas doesn't take particular care to be consistent between architectures,
the assembly portions are also not uniform and the build script needs to know
particulars of the architecture's assembly.
Instead of doing that, after this commit we first gzip each included file, then
we pack the gzipped files into a tar archive, and then we include the archive
into the binary using objcopy.
Linker script provides symbols for the start and end of the archive,
and the payload is in a self-describing format, so there is no need for any
generated code.
Note that we are doing the opposite of the conventional .tar.gz format.
It would be somewhat inconvenient to use .tar.gz since the uncompressed files
need to be aligned to page size, so we'd have to first decompress the entire
payload to determine the final position of the files (and hence the required
amount of memory).
