[CIR][ABI][Lowering] Fixes calling convention

[gitoleg]

This PR fixes two run time bugs in the calling convention pass. These bugs were found with csmith.

Case #1. Return value from a function.
Before this PR the returned value were stored in a bit casted memory location.
But for the next example it's not safe: the size of a memory slot is less than the size of return value. And the store operation cause a segfault!

#pragma pack(push)
#pragma pack(1)
 typedef struct {
    int f0 : 18;
    int f1 : 31;
    int f2 : 5;
    int f3 : 29;
    int f4 : 24;
 } PackedS;
#pragma pack(pop)

CIR type for this struct is !ty_PackedS1_ = !cir.struct<struct "PackedS1" {!cir.array<!u8i x 14>}>, i.e. it occupies 14 bytes.
Before this PR the next code

PackedS foo(void) {
     PackedS s;
     return s;
}

 void check(void) {     
     PackedS y = foo();    
}

produced the next CIR:

  %0 = cir.alloca !ty_PackedS1_, !cir.ptr<!ty_PackedS1_>, ["y", init] {alignment = 1 : i64}
  %1 = cir.call @foo() : () -> !cir.array<!u64i x 2> 
  %2 = cir.cast(bitcast, %0 : !cir.ptr<!ty_PackedS1_>), !cir.ptr<!cir.array<!u64i x 2>>
  cir.store %1, %2 : !cir.array<!u64i x 2>, !cir.ptr<!cir.array<!u64i x 2>>

As one cat see, %1 is an array of two 64-bit integers and the memory was allocated for 14 bytes only (size of struct). Hence the segfault! This PR fixes such cases and now we have a coercion through memory, which is even with the OG.

Case #2. Passing an argument from a pointer deref.
Previously for the struct types passed by value we tried to find alloca instruction in order to use it as a source for memcpy operation. But if we have pointer dereference, (in other words if we have a <!cir.ptr < !cir.ptr ... > > as alloca result) we don't need to search for the address of the location where this pointer stored - instead we're interested in the pointer itself. And it's a general approach - instead of trying to find an alloca instruction we need to find a first pointer on the way - that will be an address we meed to use for the memcpy source.

I combined these two cases into a single PR since there are only few changes actually. But I can split in two if you'd prefer

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[bcardosolopes]

Great!