Use safe API to access Tensor data pointer

[Stonepia]

The un-templated data_ptr() does not check about the type check and storage_initialized() check. Thus, it may introduce potential bugs.

By using the template data_ptr, this could throw an error like below:

torch._dynamo.exc.TorchRuntimeError: Failed running call_function torchvision.roi_align(*(FakeTensor(..., device='xpu:0', size=(1, 1024, 50, 75), dtype=torch.bfloat16), FakeTensor(..., device='xpu:0', size=(1000, 5), dtype=torch.bfloat16), 0.0625, 14, 14, 0, True), **{}):
The tensor has a non-zero number of elements, but its data is not allocated yet. Caffe2 uses a lazy allocation, so you will need to call mutable_data() or raw_mutable_data() to actually allocate memory.

[Stonepia]

Note that currently, there is no bug found in stock pytorch about this mentioned unsafe data ptr (but we found on in IPEX).

This PR doesn't address the static_cast data_ptr. These occurrences do not use the template data_ptr. Similarly, the data_ptr can't be set as const_data_ptr.

static_cast<TLMetaForAddressScalar<scalar_vals_t, depth>*>(
          addressStorage.data_ptr());

// Could not be changed to int64_t
static_cast<TLMetaForAddressScalar<scalar_vals_t, depth>*>(
          addressStorage.mutable_data_ptr<int64_t>());

// Set as mutable_data_ptr. Could not use const_data_ptr
static_cast<TLMetaForAddressScalar<scalar_vals_t, depth>*>(
          addressStorage.mutable_data_ptr());

To fix this, we may need to further change these static_cast usage (to align with PyTorch implementation).