Control recompilations/specialization tighter from helion, including triton recompilations

[Chillee]

I'd like to be able to disable the triton recompilations/specialization as well from helion. For example, i don't want the triton kernel to recompile if it's 0/1 or divisible by 16.

Triton supports do_not_specialize: triton-lang/triton#5000

Ideally I'd be able to control it natively from helion in some manner, but barring that, just passing through do_not_specialize is acceptable.

[parsshar-RH]

Hi,

I would like to work on this!

Please let me know if the following approach is correct:

• In DeviceFunction.codegen_function_def() collect the indices of all TensorSizeArg, TensorStrideArg, and SymbolArgument parameters and pass them as do_not_specialize to @triton.jit()
• When there is no dynamic integer arg (static_shapes=True) the list is empty and the decorator is unchanged.

cc: @jansel @yf225

[jansel]

For tensors this may be blocked on #934