Avoid unnecessary allocation for views which store sub/upp-diagonal data for tridiagonal solver

[AbdelhadiKara]

Please mention something about the expected sizes here. Especially in places where the size is not intuitive. E.g. by default I would expect :

• AA to have size block_size x (n_diags-1),
• BB to have size block_size x n_diags,
• CC to have size block_size x (n_diags-1)

But looking at the later code it seems that this is not the convention is chosen. It is therefore important to document the expected contents and position of the additional 2*block_size elements.

It would be good to remove these extra elements (imagine how much extra data is required if we solve in 5D space) but this should be done in a later MR and is not a priority atm as we don't currently need this solver for the 5D case.