Improved Gauss kernel generation

[RunDevelopment]

Changes:

• Added debug assertions to check assumptions in debug.
• Define a minimum sigma of 0.01. This prevents division by zero.
• Removed the scaling by $1/\sqrt{2\pi\sigma^2}$. Since kernel weights are normalized, this scaling is unnecessary.
• Remove use of f32::powf and cache more divisions. The performance impact will likely be negligible, but I'd rather remove functions of unknown complexity.

Taken together, get_gaussian_kernel_1d now generates correct kernels for all $\sigma\ge 0$.

Since I only changed the behavior for edge cases (very small sigma), the behavior of our Gaussian blur remains unchanged.