Self attention with relative positional embedding

[vgrau98]

Is your feature request related to a problem? Please describe.
Enable relative positional embedding in attention block

Describe the solution you'd like
A constructor argument specifying which type (or None) of relative embedding to use

Describe alternatives you've considered
NA

Additional context
Should be useful for some network implementation like SAM (https://arxiv.org/abs/2304.02643, #6357 )
PR opened #7346 with suggestion for decomposed relative positional embedding (https://arxiv.org/abs/2112.01526)