block_impl geometry is slow

[hunse]

For large matrices (e.g. encoders with millions of neurons), the geometry planning (the for n in items loop creating all the indices, etc., in block_impl) is quite slow.

This may also be the case for other GEMV implementations; I haven't checked.