[tk] Extend type system and indexing sufficient to generate kernel functions.

[stellaraccident]

• More boilerplate for dependent types in the KernelBuffer hierarchy:
  • Adds usage (input, output, temporary)
  • Adds element_type and maps to torch.dtype
• Adds a stub IndexingContext which can only handle constant symbol bindings right now.
• Adds a vector_codegen emitter and gets it to the point that it can emit function definitions.
• Next step for the iota kernel is to add grid bindings to the signature, emit the program_id operator and finish modeling the store from local to global memory. This can probably be done without more FX level passes, which would be needed for anything more complicated.