Refactor of the kernel dispatch API

[diptorupd]

The existing kernel dispatch API in numba_dpex.compiler.py is sub-optimal. There are several problems with the code:

• Interleaved code between dispatching a kernel and the actual representation of a kernel
• Error messages are a mix of prints and ValueError and TypeError with abstruse messages.
• Improve the USMArrayType.
• the deprecated dpctl.get_current_queue call is embedded inside the API making it hard to separate out the legacy dpctl.device_context support from the compute-follows-data programming model.
• There is no check on device backend. We always assume that the device supports SPIR-V.
• Pass llvm bitcode to llvm-spirv using stdin rather than first serializing the bitcode to disk #862
• Caching of kernels uses a dictionary and can lead to a growing memory footprint of a dpex program. Instead, Numba's caching mechanism should be used to cache compiled kernel functions.
• The check for legal global ranges is commented out.
• Compiling a kernel with a specific signature always uses dpctl.get_current_queue

The code is also not properly documented nor are there API documentation generated. Unit test coverage can be improved too.

• Generate API documentation and developer guide for the kernel API.
• Update examples to use compute-follows-data for kernels.
• Generate class UML diagrams.

These issues have to be fixed before new feature development and bug fixes can be implemented for the dpex kernel API.

#804 is a WIP PR to fix these issues and make the API modular and properly documented with high coverage.

[diptorupd]

Most things are now complete and we have separate issues to track the left over work.