Add NonEmpty variants of inits and tails

[clyring]

See #552.

The lazy versions use new implementations:

• Lazy tails got about 10% faster. (A happy accident!)
• Lazy inits got much faster:
  • For the first few chunks it is about 50% faster, due to better fusion.
  • When there are many chunks it is about 2.5x faster.

The performance of the strict versions is unchanged. The inits benchmark with small chunks is on the slow side, taking a little over 100ms on my machine.

[Bodigrim]

Great stuff!

[clyring]

It looks like my original run of the Lazy.tails benchmark with ghc-9.4 was simply unlucky. This patch actually seems to regress runtime of that benchmark by somewhere around 5-8% with ghc-9.4.3 on my machine. (It does really improve by 10% with ghc-9.2.5.) I tried looking into it but haven't found any clarity. In fact, when I revert the changes to Lazy.tails entirely but keep the rest of the patch I still observe it getting 5-8% slower than it was before the patch. So I'm inclined to ignore this regression.

Fixing the order of composition in lazy inits made the small-chunks version of that benchmark go from 2.5x faster to 4x faster than the original code. (This is because only the wrong-way composed function needed to allocate a thunk when called.)