Use FP16 operands for UMA radial FC2

[mlazos]

Summary

This adds a default-off fp16_radial_fc2_blocks path for the second unified
UMA radial projection. Forward and frozen first-order backward use FP16 GEMM
operands with FP32 outputs, while layer normalization, activation, and
surrounding values remain FP32.

The implementation keeps nonpersistent FP16 weight caches stable across CUDA
graph replay, refreshes them after state or device changes, preserves state
dictionaries, and rejects non-FP32 base precision and Hessian inference.

Performance

With torch.compile(dynamic=True), the option improved the full endpoint from
17.01415 to 16.82399 ms with fresh caches and from 17.01992 to 16.82917 ms with
warm caches. The paired mean improvement is 0.19046 ms (1.12%).

A separate static-shape comparison improved 17.71949 to 17.55717 ms, a
0.16232 ms (0.92%) reduction. Maximum observed energy, force, and stress
differences were 0.006781, 0.000299, and 0.0000167.

Configuration: H100, TF32, UMA-S-1p2, 1,000 atoms, energy/forces/stress,
umas_fast_gpu, merge_mole=True, external_graph_gen=False, internal graph
v3, internal_graph_skin=0, compile=True, compile_dynamic_shapes=True, and
full CUDA graph replay. Steady-state execution had zero recompiles and zero
dynamic CUDA graph fallbacks.

Testing

PYTHONPATH=src /home/mlazos/.conda/envs/pytorch-3.12/bin/python -m pytest -q -c packages/fairchem-core/pyproject.toml tests/core/models/uma/nn/test_unified_radial.py tests/core/models/uma/uma_fast/test_execution_backends.py tests/core/units/mlip_unit/test_inference_settings.py -m 'not gpu'
pre-commit run --files docs/core/common_tasks/ase_calculator.md src/fairchem/core/models/uma/nn/execution_backends.py src/fairchem/core/models/uma/nn/unified_radial.py src/fairchem/core/units/mlip_unit/api/inference.py tests/core/models/uma/nn/test_unified_radial.py tests/core/models/uma/uma_fast/test_execution_backends.py tests/core/units/mlip_unit/test_inference_settings.py

The focused suite passed 21 tests, and all pre-commit hooks passed. The added
GPU coverage exercises eager and compiled VJPs plus CUDA graph replay.