Add missing @propagate_inbounds and optimized form of setindex

[dennisYatunin]

Purpose

This PR fixes two issues I discovered while working on #2417:

• When recursive unrolling was replaced with unrolled_map in Group/unrolling #2413, several @propagate_inbounds annotations where accidentally dropped, since neither unrolled_map nor the anonymous functions passed to it were annotated with @propagate_inbounds. This has led to unnecessary bounds checking during certain broadcast operations, which can be seen in the CI flame graphs for ClimaCore and ClimaAtmos.
• In Add support for nonuniform data structures #2464, I implemented several functions in terms of Base.setindex, including the performance-critical struct_field_view and struct_indices. It turns out that Base.setindex does not always inline the index value, leading to unnecessary index comparisons and bounds checks during runtime. These were also visible in the CI flame graphs, though they were not as significant as the first issue.

Content

Three new functions are added to a sub-module of Utilities: unrolled_setindex, unrolled_insert, and unrolled_map_with_inbounds. These will eventually be generalized and moved to UnrolledUtilities.jl, but for now they can be explicitly imported from ClimaCore.Utilities.Unrolled. I expect we'll need to define other similar functions in the process of refactoring DataLayouts/Operators, after which we can update the UnrolledUtilities API to include them.

---

• Code follows the style guidelines OR N/A.
• Unit tests are included OR N/A.
• Code is exercised in an integration test OR N/A.
• Documentation has been added/updated OR N/A.

[imreddyTeja]

Do you have evidence that these changes make inbounds propagate as expected? Same question for the inlining of the index

[dennisYatunin]

@imreddyTeja Here are screenshots of the "Analysis: Flamegraph Profile" job from ClimaCore's CI that illustrate the difference; the first screenshot is from the latest main branch, while the second screenshot is from this branch: