Split H5D_CONTIGUOUS_REF datasets into virtual chunks in getChunkDims

[mattjala]

getChunkDims returned the full dataset shape for every layout that isn't H5D_CHUNKED*. That's right for contiguous data that a store keeps itself, but an H5D_CONTIGUOUS_REF dataset is a byte range in another file, so treating it as one chunk means that reading any element fetches, decodes and caches the entire range. HSDS has used getChunkDims for its chunk grid since 1.0, so on HSDS 1.0.x a one-element read of a linked contiguous dataset reads the whole thing. This is exposed as a massive performance regression in many circumstances. HSDS 0.9.x split these into virtual chunks itself, and this functionality was lost during the 1.0.0 move to hdf5-json.

getChunkDims now returns getContiguousLayout's split for H5D_CONTIGUOUS_REF, using the module's CHUNK_MIN and CHUNK_MAX: virtual chunks of at most ~7.9 MB. The split depends only on the dataset's shape and type. Variable-length types, which validateLayout doesn't allow for this layout, keep the single-chunk behavior. Other layouts are unchanged.

This exposes an existing HSDS bug. HSDS's getStorBytes zero-fills short range reads instead of padding them, so the final virtual chunk of a dataset that ends at end-of-file reads back as zeros. HSDS needs that one-line fix before depending on this release.

Addresses the problem described in HDFGroup/hsds#475.