GlobalMutualInformationLoss returns NaN for constant B-spline inputs

Describe the bug

GlobalMutualInformationLoss(kernel_type="b-spline") returns a non-finite loss for valid constant prediction or target tensors. Constant tensors have zero intensity range, so the B-spline bin width becomes zero and normalization produces NaNs.

Two related boundary failures exist in the same path:

• num_bins <= 4 cannot accommodate the implementation's two-bin padding on each side, leaving a zero or negative interior denominator.
• In float16, the cubic B-spline distance polynomial can overflow before masking.

To Reproduce

import torch
from monai.losses import GlobalMutualInformationLoss

pred = torch.zeros((1, 1, 8, 8), requires_grad=True)
target = torch.ones_like(pred)
loss = GlobalMutualInformationLoss(kernel_type="b-spline")
result = loss(pred, target)
print(result, torch.isfinite(result))

On the current dev revision, result is non-finite.

Expected behavior

Valid constant tensors should produce a finite loss and finite gradients. B-spline configurations with too few bins should fail at construction with a clear ValueError.

Screenshots

Not applicable.

Environment

MONAI version: 1.6.0rc1+46.g3ee058bd
MONAI revision: 3ee058bdd16dd4a566d23d3f84687c3c35268a36
Python version: 3.13.13
PyTorch version: 2.12.0
NumPy version: 2.4.6
System: macOS 27.0 arm64
CUDA available: False

Additional context

A focused fix and regression coverage are proposed in #9022.

Start at the GlobalMutualInformationLoss entry point and reproduce the constant-tensor case from the issue. Check the B-spline path for zero ranges, too few bins, and float16 behavior, then verify finite loss and gradients, construction-time ValueError handling, and regression coverage as proposed in #9022.