[Breaking] Adding reductions across multiple axes

[coreylowman]

Adds ability to call reductions along multiple axes. 🎉

This achieves two things mainly:

1. All reductions can now be used for any amount of axes
2. Makes reductions consistent with broadcast, where the output type is used to infer axes.

Closes #163

Breaking Changes

• Remove -1 from valid axes, add trait HasLastAxis to use in generic functions instead
• Making full reduction functions generic over axes: mean, sum, max, min
• Adding fully generic versions: normalize, stddev, var
• Tensor methods for reductions can infer axes & output type similar to .broadcast().
• Removing single axis reduction functions fn *_axis():
  • mean_axis
  • sum_axis,
  • max_axis
  • min_axis
  • normalize_axis
  • std_axis
  • var_axis
• Rename HasAxis to HasAxes
• Rename Broadcast to BroadcastTo

Non breaking changes

• Adding AllAxes & removing device reduce all