Strided vector functions

[fredrik-johansson]

It would be useful to have strided version of a subset of _gr_vec functions, allowing efficient application to columns and diagonals of matrices, reversed vectors, odd-even subvectors, etc. For example one could have an assignment function like this:

_gr_vec_set_strided(res, x, xstride, y, ystride, len, ctx)

Potentially, this could replace a number of specialized vector functions, such as those mixing vectors and scalars. For example, instead of

_gr_vec_add_scalar(res, x, len, y, ctx)

one could do

_gr_vec_add_strided(res, x, 1, y, 0, len, ctx)

Likewise for _gr_vec_dot / _gr_vec_dot_rev. Note that a _gr_vec_dot_strided could be used for basecase matrix multiplication without the need for a temporary transpose (though the temporary transpose is better for cache efficiency with large matrices anyway). Likewise for other accumulating operations like sum, max. Currently arb_dot, acb_dot and _fmpz_vec_dot_general have this functionality.

For non-accumulating (entrywise) operations, we mostly care about this for machine types and some near-machine types like fmpz where the overhead of calling elementwise operations in a loop can be noticeable.

Unfortunately, strided functions may not be able to replace non-strided ones entirely for such types: there are more parameters to pass around making function calls slightly slower, and some loops will run slower with a runtime offset instead of a fixed compile-time offset like 1 or -1. One can have branches for special strides, but then the branches also have an O(1) penalty.

[fredrik-johansson]

Related: _nmod_vec_dot_ptr is basically obsolete after #2162 and could be supplanted by a strided version.