Encode expected/actual info in ShapeError

[bluss]

Add a (limited) way to add specific information to a ShapeError

Admittedly wonky, but maybe worthwhile.

Result<(), ShapeError> used to be 1 byte, and with this change it
expands to 16 bytes (2 usize on 64-bit).

The remaining 15 bytes are used for optimistically packing as much of
extra info into the error message as possible.

For example we can store expected/actual index for errors (for example
index out of bounds or axis out of bounds, these are not so commonly
handled with ShapeError).

With this change it is supported:

• Expected/actual index with 7 bytes per index
• Expected/actual shape with 7-14 bytes per shape
  supports storing shapes with one or two bytes (< 256²) per dimension,
  with limited ndim.

[bluss]

I guess I needed to experiment with this no-alloc approach. This is a good starting point for also experimenting with an error in a Box, which also does fine on microbenchmarks. Maybe it can be optional.

[bluss]

After exploring this, I'm ready to try adding something like ShapeError(Box<ErrorDetails>) to pack in more information. Thoughts? What I don't like is the code bloat that results from cleanup pads (code to drop the extra box in various places). It could even be made optional, to have such error details.