List of critical bugs giving incorrect results without error messages

Incorrect results will be given in these versions and situations with no error messages:

* v3.0.0~v3.2.0b0: in the PyTorch backend with LAMMPS, the compressed `se_e2_a` model may produced incorrect results; reported by #5092, fixed by #5524
* v3.0.0~v3.0.2: when using DeepEval to evaluate DPLR models, long-range term is not added; reported by #4625, fixed by #4753.
* v2.2.5~v2.2.7: random incorrect GPU results due to invalid GPU memory write (see #3150, #3102, #3094, #3034, #3242); fixed by #3157
* ~v2.2.6: the compressed model generates incorrect results when `neuron[0]` is 1; fixed by #2933
* ~v2.2.6: the compressed `se_e2_a` model generates incorrect results for symmetrical coordinates when `sel` is just enough (this condition is rare as most structures during simulations are not symmetrical); fixed by #2924
* v2.2.4~v2.2.5: `se_atten_v2` (or `se_atten` with `smooth_type_embdding=True`) model gives inconsistent energy and force; fixed by #2905
* v2.2.3~v2.2.5: incorrect results when compressed training `se_atten` model; reported by #2865, fixed by #2898
* v2.2.0~v2.2.4: incorrect results from DPLR training; reported by #2862, fixed by #2864
* ~v2.2.3 with TF >=2.13: incorrect results on GPUs; reported by #2660, fixed by #2661
* v2.1.0~v2.2.1: incorrect results from `se_e3` compressed model; reported by #2250, fixed by #2552
* v2.2.0: frozen models can not be performed on LAMMPS; reported by #2398, fixed by #2390
* v2.1.0~v2.1.4: incorrect force results on CPUs when rcut * 2 > box size; fixed by #1862
* ~v2.1.0 with LAMMPS >=30Jul2021: incorrect LAMMPS `pppm_dplr` results; fixed by #1637
* ~v2.0.3 with LAMMPS >=30Jul2021: incorrect LAMMPS `compute deeptensor/atom` results; reported by #1381, fixed by #1382
* ~v2.0.1 with LAMMPS >=30Jul2021: incorrect LAMMPS `pair deepmd` results; reported by #1109, fixed by #1128

----
The potential energy is not continuous in the following cases:
* In all versions, `loc_frame` model
* In all versions, `se_atten` model with `set_davg_zero=False` (default=False in v2.1.5 and True in >=v2.2.0) and `smooth_type_embdding=False` (added in v2.2.4, default=False)
* ~v2.2.5, `se_atten` or `se_atten_v2` model with non-zero `attn_layer`; reported by #2842, fixed by #2905