[InMemoryDataset redesign] Optimize fancy indices

[crusaderky]

• Part of [DNM] Redesign InMemoryDataset data model (libhdf5 direct access) #386
• Blocked by and incorporates [InMemoryDataset redesign] Object-oriented as_subchunk_map #372
• Blocked by and incorporates [InMemoryDataset redesign] New function asarray() #384
• Incorporates [InMemoryDataset redesign] ix_with_slices, format_ndindex #387 (no dependency; it's just to keep branching at a minimum)

This PR is made of several commits; the first commits are the incorporated PRs listed above. See only final commit for review.

PR contents

• Simplify fancy indices down to slices whenever possible - first try to do so globally, and if that's not possible try again locally for each chunk. This will produce an immediate benefit even in the legacy as_subchunk_map implementation, as it avoids an expensive intermediate deep-copy.
• Speed up monotonic ascending fancy indices from O(n^2) to O(n*logn)