Remove PoW registration, refactor register() to use register_limit

[basfroman]

Includes:

• Remove PoW registration extrinsic (sync and async) and pallet method
• Remove async_pow.py, register_cuda.py, pow.py from bittensor/utils/registration/
• Extract torch compatibility utilities into torch_utils.py
• Refactor Subtensor.register() and AsyncSubtensor.register() to use register_limit_extrinsic under the hood
• Auto-calculate limit_price as recycle * 1005 // 1000 (0.5% tolerance) when not provided
• Dispatch to root_register_extrinsic when netuid == 0
• Accept limit_price: Optional[Balance] parameter with check_balance_amount validation
• Remove PoW unit tests (sync and async), add tests for new register() logic
• Add E2E tests for register() with auto-calculated limit_price
• Update import paths from pow to torch_utils in tests

Base Subtensor PR

• Neuron Registration opentensor/subtensor#2382