[SPIR-V][Doc] add draft SPIR-V extension for 16-bit atomics

[bashbaug]

Adds the proposed SPV_INTEL_shader_atomic_bfloat16 extension specification, which extends the floating-point atomic support in SPIR-V to support floating-point atomic addition, minimum, and maximum on 16-bit bfloat16 floating-point numbers in memory.

[bashbaug]

We may want to consider broadening this extension to support 16-bit atomics more generally vs. just bfloat16 atomics. Here's what this might look like:

• Add a capability for 16-bit integer atomics, perhaps Int16Atomics, similar to the Int64Atomics capability for 64-bit integer atomics. We could slice this more finely if we would like and add capabilities for specific 16-bit atomic operations.
• Figure out a more appropriate name for the extension, since it wouldn't just be for bfloat16 atomics anymore.

Just like in this current draft for bfloat16 atomics, I don't think we would need any new instructions, and would just support 16-bit types for existing atomic instructions.

[bashbaug]

I've generalized this extension for 16-bit atomics, as described above. The SPIR-V extension still contains separate capabilities, so it is possible to only support bfloat16 atomics if desired, but it is also possible to support 16-bit integer atomics as well. Note that fp16 atomics are already supported by other extensions.

The OpenCL extension proposal in this PR shows possible API queries so implementations could advertise only the 16-bit atomic operations they support.