Various codegen fixes

Here we fix a number of backend bugs to do with sub-word arithmetic, many of which involving signedness:

• a bug in code generation for shifts on x86 (#20626 (closed))
• some bugs in Cmm constant folding (#20634)
• warn in Cmm lint on large shifts which are undefined behavior (#20637)
• incorrect sign extension during loads in the AArch64 NCG (#20638)
• incorrect signedness in arithmetic in the AArch64 NCG (#20644 (closed))
• incorrect handling of pext and pdep by the LLVM backend (#20645)
• a pair of bugs in the C backend (#20745)

This should allow bgamari/test-primops> to pass on all platforms.

TODO:

• add tests