cl_khr_fp16 is assumed to work on all hardware, even when this is not the case

[softcookiepp]

When using certain older GPUs via the Rusticl backend (I am using a RX 570 for this,) the following error shows up when trying to perform any operations on float16 tensors:

`RuntimeError: OpenCL Compile Error

input.cl:1:26: warning: unsupported OpenCL extension 'cl_khr_fp16' - ignoring [-Wignored-pragmas]
input.cl:3:3: error: use of undeclared identifier 'half4'
input.cl:4:15: error: unknown type name 'half4'
input.cl:4:36: error: use of undeclared identifier 'half4'
input.cl:4:44: error: use of undeclared identifier 'val0'
input.cl:4:51: error: use of undeclared identifier 'val0'
input.cl:4:60: error: use of undeclared identifier 'val0'
input.cl:4:67: error: use of undeclared identifier 'val0'
input.cl:4:76: error: use of undeclared identifier 'val0'
input.cl:4:83: error: use of undeclared identifier 'val0'
input.cl:4:92: error: use of undeclared identifier 'val0'
input.cl:4:99: error: use of undeclared identifier 'val0'
Error executing LLVM compilation action.`

Here is the code that can be used to reproduce this bug provided you have the hardware:

`from tinygrad import Tensor, dtypes

t = Tensor([1, 2, 3, 4], dtype = dtypes.float16)
(t + t).realize()`

Given that cl_khr_fp16 is an optional OpenCL extension, it may be better to not assume that it exists on all backends, and to use type promotion or similar means of eliminating the issue.

[chenyuxyz]

implicitly upcasting to float32 if half is not available is bad, but maybe we can support a global flag to say all half should be interpret as float