Clarify use of cross-correlation for FIR filters

[glopesdev]

The underlying implementation of the FirFilter operator uses the filter2D function from OpenCV to efficiently apply the filter.

Although the documentation of filter2D itself states that it "convolves an image with a kernel", in fact the implementation actually computes the cross-correlation. This is a surprisingly common, if misleading, assumption in the computer vision and image processing field, likely coming from the use of either symmetric kernels (in which case correlation is equal to convolution), or learned kernels (in which case the learning process can easily deal with reversed kernel weights).

Nevertheless, the current reference documentation of the operator can be misleading if one is attempting to use FirFilter with externally computed asymmetric kernels for temporal convolution. This PR adds a new remarks section clarifying the relationship between correlation and convolution, and how the kernel must be processed to obtain the correct result in case of an asymmetric kernel.

Fixes #1761