DX12: give each shader entry point its own intermediate files - #20107

Summary

Fixes three problems in how a shader's entry points are compiled. All three are present when stages
compile one after another, which is how the builder runs them today, and this PR keeps it that way.

1. DX12 intermediates collided across entry points

CompilePlatformInternal derived every intermediate file name from the source file alone. Every entry
point of a shader compiles from the same source into the same temp folder, so each stage wrote the same
files:

• <stem>.dxil.bin and <stem>.dxil.txt
• the prepended source
• with specialization constants, <stem>.dxil.patched.bin and <stem>.offsets.json

Each stage overwrote what the previous one wrote, so only one stage's intermediates survived. It also
means a tool calling the DX12 interface directly cannot compile two stages of one shader at the same
time, because they write to the same paths.

Intermediate files are now named per entry point (<stem>.<entry>.dxil.bin and so on), and the prepended
source gets the same suffix. RayTracing has no entry point and uses the profile name instead, the same
way the .pdb name was already made unique.

2. Only the last stage's byproducts reached the caller

ShaderVariantAssetBuilder called outputByproducts.emplace() once per entry point. outputByproducts
is an AZStd::optional, and emplace on an optional replaces its current value, so every stage
discarded the previous stage's intermediate paths. The byproducts are now merged.

3. Entry point order depended on hash order

Entry points were iterated in unordered_map order. Which stage ran last, and therefore which dynamic
branch count was reported, depended on the hash rather than on the shader. Entry points are now compiled
in name order.

Not in this PR: concurrent stage compilation

An earlier revision compiled entry points in parallel in ShaderVariantAssetBuilder. That code is
shared by every backend, and only DX12 was made safe:

• Vulkan names <stem>.spirv.bin, <stem>.spirv.txt and its prepended source from the source file
  alone, so its stages collide in the same way DX12's did.
• Metal has the same file name collision (.spirv, .metal, .air, .metallib). It also keeps
  per-compile state on its ShaderPlatformInterface (m_srgLayouts, m_argBufferEntries), which is
  written during compilation.
• RHI::PrependFile lazily initializes a function-local static without synchronization. This is
  shared by all backends, including DX12.

Concurrent stage compilation can come back in its own PR once every backend is safe.

Testing

• DX12: shaders with vertex and pixel entry points process, and the intermediate files in the temp
  folder are named per entry point.
• Specialization constants: the per-stage .dxil.patched.bin and .offsets.json are produced and the
  shader renders correctly.
• Vulkan and Metal: unchanged by this PR apart from the entry-point ordering and byproduct merge in the
  shared builder.