🐛[BUG]: Threshold is actually optional if using custom_op

Version

0.1.0

On which installation method(s) does this occur?

Source

Describe the issue

The ConvergenceTreshold requires a threshold even if it's not needed used because we're using a custom operation.

Should we either make it optional if custom_op is set or pass it as an arg to the custom_op? I'd suggest the latter so users can tweak the tolerance of a convergence hook more easily.

Minimum reproducible example

import torch
from nvalchemi.dynamics import ConvergenceHook


def _custom_noop(x: torch.Tensor) -> torch.Tensor:
    return torch.ones(1, dtype=torch.bool, device=x.device)

ConvergenceHook = ConvergenceHook(
    criteria={"key": "custom_metric", "custom_op": _custom_noop},
)

Relevant log output

Traceback (most recent call last):
  File "/mnt/c/Users/lward/Code/nv-open/nv-alchemi/nvalchemi-toolkit/scripts/thr-mre.py", line 8, in <module>
    ConvergenceHook = ConvergenceHook(
        criteria={"key": "custom_metric", "custom_op": _custom_noop},
    )
  File "/mnt/c/Users/lward/Code/nv-open/nv-alchemi/nvalchemi-toolkit/nvalchemi/dynamics/base.py", line 2288, in __init__
    self.criteria = [_ConvergenceCriterion(**criteria)]
                     ~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^
  File "/mnt/c/Users/lward/Code/nv-open/nv-alchemi/nvalchemi-toolkit/.venv/lib/python3.13/site-packages/pydantic/main.py", line 250, in __init__
    validated_self = self.__pydantic_validator__.validate_python(data, self_instance=self)
pydantic_core._pydantic_core.ValidationError: 1 validation error for _ConvergenceCriterion
threshold
  Field required [type=missing, input_value={'key': 'custom_metric', ...noop at 0x7a7d06ee7100>}, input_type=dict]
    For further information visit https://errors.pydantic.dev/2.12/v/missing

Environment details

Start in nvalchemi/dynamics/base.py at _ConvergenceCriterion and ConvergenceHook, then reproduce the provided example with the custom operation. Trace how threshold validation interacts with custom_op and determine which behavior the issue should adopt. Done means a custom_op criterion can be constructed as intended and regression coverage verifies the selected threshold behavior.