[SPIR-V] Inefficient codegen when writing to RWByteAddressBuffer

Description
When writing/reading to/from a RWByteAddressBuffer, it emits a lot of loads/writes using a uint each.
Even though the driver might catch this and fix it at PSO creation time certain validation tools might still see performance degradations.
In this issue: https://github.com/KhronosGroup/Vulkan-ValidationLayers/issues/9317#issuecomment-2616396801 compilation takes a loooong time because every access needs to be validated individually.
If there is a way to reduce this, for example by using bitcasts as suggested by devshgraphicsprogramming in a different issue (https://github.com/microsoft/DirectXShaderCompiler/issues/7038#issuecomment-2572959501) it could help a lot to reduce overhead with validation layers and potentially other tools as well. Maybe also drivers that are written less well can also benefit from this.

Steps to Reproduce
RWByteAddressBuffer or ByteAddressBuffer load/write with a relatively big struct. This will emit a lot of bloat https://godbolt.org/z/qE95f8jvb. Where each index needs to be bounds checked by the tool.

Actual Behavior
Slowdowns on validation tools.

Environment

• DXC version: Latest from a few days back
• Host Operating System: N/A (though the issue happens on Android)

Start with the Godbolt reproduction and inspect the generated SPIR-V for RWByteAddressBuffer accesses to understand the repeated loads, writes, and bounds checks. Compare output for a relatively large struct before and after any investigation, and treat reduced validation overhead without changing buffer access behavior as the completion criteria.