Fix minor heap stats for bigarrays allocated with custom data pointers.

[toots]

While debugging a user-reported memory leak in liquidsoap here: savonet/liquidsoap#2054, after going down the rabbit hole, I ended up looking at this piece of code:

  data = aligned_alloc(alignment, len);
  if (data == NULL)
    uerror("aligned_alloc", Nothing);
  ans = caml_ba_alloc_dims(CAML_BA_MANAGED | CAML_BA_C_LAYOUT | CAML_BA_UINT8,
                           1, data, len);

I was able to confirm that, if data was passed as NULL here, the compiler would allocate the memory itself and no memory leak would be happening. Furthermore, allocating data using malloc like the compiler does also lead to a memleak. I was also able to confirm that, by forcing a Gc.full_major() on each allocation, no memory leak was happening.

After more investigations and head scratching, it turned out that the problem comes from this code:

  size = 0;
  if (data == NULL) {
    num_elts = 1;
    for (i = 0; i < num_dims; i++) {
      if (caml_umul_overflow(num_elts, dimcopy[i], &num_elts))
        caml_raise_out_of_memory();
    }
    if (caml_umul_overflow(num_elts,
                           caml_ba_element_size[flags & CAML_BA_KIND_MASK],
                           &size))
      caml_raise_out_of_memory();
    data = malloc(size);
    if (data == NULL && size != 0) caml_raise_out_of_memory();
    flags |= CAML_BA_MANAGED;
  }
  asize = SIZEOF_BA_ARRAY + num_dims * sizeof(intnat);
  res = caml_alloc_custom_mem(&caml_ba_ops, asize, size);

The size variable in this code is passed as 0 when using a custom data pointer, thus confusing the Gc about how much data the allocated value actually represents.

The natural fix seems to be to make sure that size is always filled.

[damiendoligez]

This is almost exactly the same patch as #11022, but this PR doesn't take the CAML_BA_MANAGED flag into account.