[naga wgsl-in] Ensure constant evaluation correctly handles Composes of vector ZeroValues

[jamienicol]

Connections
Fixes #7137

Description
Constant evaluation often relies on the expressions being evaluated being either a Literal, or a vector Compose for which it can iterate over each component.

Currently this is achieved by calling two functions:

• eval_zero_value_and_splat() which transforms a scalar ZeroValue into a Literal, or a Splat or vector ZeroValue into a Compose of Literals.

• proc::flatten_compose() takes potentially nested nested Compose and Splat expressions and produces an flat iterator that yields each component. eg vec3(vec2(0), 0) would yield Literal(0), Literal(0), Literal(0).

For component-wise vector operations, we can then iterate through each component of the flattened compose and apply the operation. When there are multiple operands it is crucial they have both been flattened correctly so that we use the corresponding component from each operand together.

Where this falls short is if a vector ZeroValue is nested within a Compose. eg vec3(vec2(), 0). flatten_compose() is unable to flatten this, and the resulting iterator will yield ZeroValue, Literal(0)

This causes various issues. Take binary_op(), for example. If we attempt to add vec3(1, 2, 3) to our unflattenable vec3(vec2(), 0) this should be evaluated component-wise as 0 + 1, 0 + 2, and 0 + 3. As this has not been correctly flattened, however, we will evaluate vec2() + 1, and 0 + 2, which is simply incorrect.

To solve this, we make eval_zero_value_and_splat() recursively call itself for each component if the expression is a Compose. This ensures no ZeroValues will be present during flatten_compose(), meaning it will successfully fully flatten the expression.

Testing
Added snapshot tests

Checklist

• Run cargo fmt.
• Run cargo clippy. If applicable, add:
• Run cargo xtask test to run tests.
• Add change to CHANGELOG.md. See simple instructions inside file.