Fix ccall return value boxing on ARM/AArch64

[yuyichao]

We previously rely on the extra allocation from the GC to keep the stores inbounds.
This is broken by the allocation optimization since the stack allocation will only have
the requested bytes and not more.

[vtjnash]

Seems like this should be using our llvm_type_rewrite so it doesn't inhibit llvm optimizations.

[yuyichao]

We are only storing the value so as long as the size is smaller or the same we don't care what type it actually is so unconditionally going through llvm_type_rewrite is unnecessary and can generate extra code.

As long as we are doing this only when there's a size mismatch (i.e. what this PR is doing), it is exactly the same as llvm_type_rewrite. I didn't use it since most of the useful logic from the function is already here to check the size....

[yuyichao]

Actually I missed the trunc case in llvm_type_rewrite so I guess I can do that. It doesn't actually matter for the real code though since the type we have here always have an aggregate....

[vtjnash]

Ah, didn't realize this was just for structs. (In which case, I think llvm_type_rewrite should be doing the same as this code?) Anyways, this is probably fine, was just considering opportunities for consolidating code.

[yuyichao]

FWIW, I think this can share code better with memcpy (the only use of the value is storing anyway). I didn't do it since I don't want to have conflict with #23352 but maybe I should just rebase on it.....