[Misc] add store intf with tensor addr ptr

[mag1c-h]

Purpose

The C++ base class in Store already takes a raw pointer, now add the same pointer-based API to the Python base class to keep it extension-friendly.

Modifications

Add new pointer-based API to the Store Python base class.

Test

CI passed with new added/existing test.

[pyxyzc]

@mag1c-h The semantics of the offset and size params kind of overlap here. Host–Storage transfers are already defined to operate on a page/block basis. Plz consider whether it really makes sense to expose this flexibility to the upper inference engines.

[mag1c-h]

@mag1c-h The semantics of the offset and size params kind of overlap here. Host–Storage transfers are already defined to operate on a page/block basis. Plz consider whether it really makes sense to expose this flexibility to the upper inference engines.

@pyxyzc For the following reasons, we need to expose both offset and size parameters in the interface:

1. Block semantics differ across storage layers:
   • On Host/Device, a "block" refers to a layer-block.
   • On Storage, a "block" contains all layers, each mapped to a different offset.
2. GQA partitioning introduces per-tp offsets:
   When tp_range > 1 in GQA models, each tp starts from a distinct offset.
3. Future-proofing for variable-size metadata:
   We may store profiling/characterization data where a single batch could contain entries of differing sizes.

[ygwpz]

do we really need two api here?

def load(self, block_ids: List[str], offset: List[int], dst_tensor: List[torch.Tensor]) -> Task
def fetch_data(self, block_ids: List[str], offset: List[int], dst_addr: List[int], size: List[int],) -> Task

i think fetch data is what we need final

[mag1c-h]

do we really need two api here?

def load(self, block_ids: List[str], offset: List[int], dst_tensor: List[torch.Tensor]) -> Task
def fetch_data(self, block_ids: List[str], offset: List[int], dst_addr: List[int], size: List[int],) -> Task

i think fetch data is what we need final

They share the same impl, but keeping both API saves every caller a rewrite.