[NPU] align fused linear JSD forward with accum_dtype API

[zheliuyu]

Summary

Upstream #1322 added accum_dtype to fused linear JSD (LigerFusedLinearJSD, LigerFusedLinearJSDFunction.apply). On NPU, liger_kernel.ops resolves to the Ascend backend, whose LigerFusedLinearJSDFunction.forward still used the old signature. All test/transformers/test_fused_linear_jsd.py cases failed with:

TypeError: forward() takes from 5 to 9 positional arguments but 10 were given

This PR updates backends/_ascend/ops/fused_linear_jsd.py to accept accum_dtype, pass it through fused_linear_jsd_forward, and mirror the default path: optional fp32 grad_weight accumulation with cast back to weight dtype. backward return tuple is extended by one None for the new argument.

Testing Done

• Hardware Type: All NPUs
• run make test to ensure correctness
• run make checkstyle to ensure code style
• run make test-convergence to ensure convergence

[zheliuyu]

@Tcc0403 This PR is ready for review. Thanks.

[Tcc0403]

LGTM