Add missing atomic operators, refactor atomic operators, move atomic operators to detail namespace.

[bdice]

Description

This PR does a thorough refactoring of device_atomics.cuh.

• I moved all atomic-related functions to cudf::detail:: (making this an API-breaking change, but most likely a low-impact break)
• I added all missing operators for natively supported types to atomicAdd, atomicMin, atomicMax, etc. as discussed in [BUG] Sum and multiply aggregations promote unsigned input types to a signed output #10149 and Revert sum/product aggregation to always produce int64_t type #14907.
  • This should prevent fallback to the atomicCAS path for types that are natively supported for those atomic operators, which we suspect as the root cause of the performance regression in [BUG] Performance regression in cuDF after #14679 #14886.
• I kept atomicAdd rather than cudf::detail::atomic_add in locations where a native CUDA overload exists, and the same for min/max/CAS operations. Aggregations are the only place where we use the special overloads. We were previously calling the native CUDA function rather than our special overloads in many cases, so I retained the previous behavior. This avoids including the additional headers that implement an unnecessary level of wrapping for natively supported overloads.
• I enabled native 2-byte CAS operations (on unsigned short int) that eliminate the do-while loop and extra alignment-checking logic
  • The CUDA docs don't state this, but some forum posts claim this is only supported by compute capability 7.0+. We now have 7.0 as a lower bound for RAPIDS so I'm not concerned by this as long as builds/tests pass.
• I improved/cleaned the documentation and moved around some code so that the operators were in a logical order.
• I assessed the existing tests and it looks like all the types are being covered. I'm not sure if there is a good way to enforce that certain types (like uint64_t) are passing through native atomicAdd calls.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[karthikeyann]

I changed all uses of atomicAdd to use cudf::detail::atomic_add . I'd like to retain the previous behavior

I support to retain previous behaviour. If not, it introduces another templated header into include dependency of other files. Templated atomicAdd is created to support unsupported types in CUDA. It's not required for these other files.

[GregoryKimball]

@SurajAralihalli FYI

[GregoryKimball]

@SurajAralihalli Would you like to share a review on this change?

[bdice]

/merge