MLX Elementwise and scalar dispatches

[jessegrabowski]

Description

Some simple/boring stuff for the MLX backend:

• Refactor scalar dispatches to copy the JAX template. We use nfunc_name to look in the mx namespace and get everything we can. There's a dictionary of overrides for stuff that doesn't follow the array standard (invert is bitwise_invert and true_divide is divide
• Move scalar stuff to scalar.py, leave CARReduce and Elementwise in elementwise.py
• Add second and identity dispatches
• Add Composite dispatch

Related Issue

• Closes #
• Related to #

Checklist

• Checked that the pre-commit linting/style checks pass
• Included tests that prove the fix is effective or that the new feature works
• Added necessary documentation (docstrings and/or example notebooks)
• If you are a pro: each commit corresponds to a relevant logical change

Type of change

• New feature / enhancement
• Bug fix
• Documentation
• Maintenance
• Other (please specify):

[jessegrabowski]

In jax OpFromGraph dispatch we're calling the jax rewrites on the inner graph, but not in Composite. I followed that lead, but it seems inconsistent

[ricardoV94]

Composite is never optimized internally, it's supposed to already be a good graph. OpFromGraph and Scan are, but we should actually explicitly optimize as a rewrite, not at compile/runtime. We are working towards that... but it takes time

Which is why I want to make their inner graphs frozen as well