ET: Implement ET backend

[vidas]

This is an umbrella issue to track implementation of ET accelerator backend for tinygrad.

There are three main areas that need to be implemented:

1. Runtime:

• Device management - maps directly to etrt.
• Allocator - maps directly to etrt.
• Command queue - effectively maps to etrt streams, but there may be lifetime issues and we can run one kernel at a time, probably should start with synchronous implementation.
• Program - maps directly to etrt kernel.

2. Compiler - will use custom GNU toolchain.

• Q. Where do we put all the build requirements, like linker.ld?
• Q. Do we include unsupported instruction checker? As part of tests?

3. Renderer. The most difficult part. ET runs kernels written in C, but there are enough architectural peculiarities to make it pretty complicated - it's neither "CPU backend", nor very GPU like (pretty close though).

Based on project architecture, ET related changes will likely touch multiple places of the code tree, including some "generic" modules - optimizer, beam search, etc. tinygrad is very fast moving project, so our goal should be to upstream ET related changes as soon as possible and until then - keep et branch in sync with upstream. This may prove a challenge.

Obviously, all tinygrad development guidelines apply.