mindtorch_v2: unify 43 backward ops to device-agnostic redispatch

[lvyufeng]

Summary

• Eliminate 43 device.type == "npu" guards from backward functions in autograd.py, converting to device-agnostic redispatch composites
• 22 A-YES ops (relu, sigmoid, softmax, etc.): delete ACLNN guard, keep existing redispatch code
• 21 B ops (diagonal, tensordot, cdist, etc.): delete ACLNN guard and numpy fallback, keep only redispatch composite
• Follows PyTorch architecture: backward formulas route through dispatch system automatically — no manual device branching
• Fix CPU reshape to support -1 dimension inference (matching PyTorch behavior)
• Fix 6 redispatch("to", ...) calls missing required device positional argument
• Net reduction of ~675 lines (deleted 1095, added 420)
• Remaining 20 device guards are Phase 2 (conv, pool2d/3d, upsample, grid_sample — need ACLNN large kernels)

Test plan

• pylint-check passes
• mindtorch-v2-test passes

🤖 Generated with Claude Code