[Distributed][HCCL] Validate all_to_all_single split sums against tensor numel

[lvyufeng]

Summary

• Add explicit all_to_all_single validation to ensure split-size sums match input/output tensor numel before collective dispatch.
• Keep existing HCCL matrix split-compatibility checks, but fail earlier with clearer user-facing errors for malformed split/numel combinations.
• Add HCCL multicard tests covering 2/4/8 cards for both input-sum mismatch and output-sum mismatch paths.

Key Changes

• src/mindtorch_v2/distributed/__init__.py
• tests/mindtorch_v2/test_hccl_all_to_all_single_split_numel_validation_multicard.py

Test Plan

• PYTHONPATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:src pytest -q tests/mindtorch_v2/test_hccl_all_to_all_single_split_numel_validation_multicard.py
• PYTHONPATH=/usr/local/Ascend/ascend-toolkit/latest/python/site-packages:src pytest -q tests/mindtorch_v2/test_hccl_all_to_all_single_invalid_splits_multicard.py tests/mindtorch_v2/test_hccl_all_to_all_single_async_unequal_multicard.py
• PYTHONPATH=src pytest -q tests/mindtorch_v2/contract/test_schema_registration_order.py tests/mindtorch_v2/contract/test_schema_coverage.py

Notes

• HCCL multicard checks were validated on local Ascend hardware; CI without Ascend devices may not exercise these paths.