fix(loss): clarify intensive normalization docs and add test coverage

[Copilot]

Review feedback identified that intensive parameter documentation overstated its scope—implying 1/N² vs 1/N applies universally when it only affects non-Huber MSE energy/virial terms.

Documentation Updates

Updated docstrings across all backends to accurately describe behavior:

• MSE energy/virial: 1/N² when intensive=True, 1/N otherwise
• MAE: remains 1/N regardless of intensive
• Huber: residuals already normalized by 1/N before Huber formula, so flag has limited effect

Files: deepmd/{utils/argcheck.py, tf/loss/ener.py, pd/loss/ener.py, dpmodel/loss/ener.py, pt/loss/ener.py, pt/loss/ener_spin.py, dpmodel/loss/ener_spin.py}

Comment Fix

Changed misleading "RMSE normalization exponent" → "loss normalization exponent" in TF backend.

Test Coverage

• Added intensive parameter to cross-backend consistency test parameterization
• Added TestIntensiveNatomsScaling regression tests verifying:
  • MSE loss doesn't scale linearly with N when intensive=True
  • intensive=True vs False produces expected 1/N ratio difference