Metal backend features

[tongzhou80]

Ranked by priority:

• Support passing scalars to kernel, including Python and NumPy scalars
• Support 2D array indexing

Probably won't do:

• Support range with > 1 args
• Support nested prange
• Support more data types

Not planned:

• Support atomic updates

Advanced:

• Use float4 or float2 (auto vectorize), not sure if this is gonna be helpful

[tongzhou80]

Support passing scalars to kernel, including Python and NumPy scalars
Done