[FEA]: Introduce cache-modified input iterator into cuda.parallel

[gevtushenko]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

Not sure

Is your feature request related to a problem? Please describe.

Usage of cuda.parallel in applications like llm.c (example) is currently blocked by lack of cache-modified iterators support.

Describe the solution you'd like

We need an functional alternative of cache-modified iterator in cuda.parallel.itertools. Design might follow the API that @fbusato came up with in #2487. For instance:

d_input = cp.array([8, 6, 7, 5, 3, 0, 9], dtype=dtype)
d_streaming_input = cudax.itertools.accessor(d_input, "eviction_policy::no_allocation")
cudax.reduce(d_streaming_input)

should lead to streaming loads of d_input (ld.global.cs instruction in PTX)

Describe alternatives you've considered

No response

Additional context

No response