size b2nd shape buffers to B2ND_MAX_DIM in ndmean and ndlz

[alhudz]

Repro: store a b2nd array whose metalayer declares ndim in 9..16, with the ndmean filter (or the ndlz codec), then read it back. ASAN reports a heap write past the shape/chunkshape/blockshape buffers in b2nd_deserialize_meta.
Cause: ndmean_forward/ndmean_backward and ndlz4_compress/ndlz8_compress allocate those buffers with 8 slots, but b2nd_deserialize_meta fills up to B2ND_MAX_DIM (16) entries straight from the metalayer. ndmean then indexes its NDMEAN_MAX_DIM stack arrays with the same unchecked ndim.
Fix: size the buffers to B2ND_MAX_DIM, as ndcell and the zfp codec already do, and reject ndim outside 1..NDMEAN_MAX_DIM in ndmean the way ndcell does.

[FrancescAlted]

Already addressed in #781.