* Impl `Hash` for `msl::Options`

* Remove ordered float

* [msl] derive everything on Options manually

* [msl] switch from Arena to Vec for the options

* Fix the clippy error by allowing derive_hash_xor_eq

* Fix skybox test

Co-authored-by: Dzmitry Malyshau <kvarkus@gmail.com>

Add a Option<u16> to the snapshot `Params` to let snapshots request specific
Desktop GLSL versions. The default is GLSL ES 3.10.

It would be more general to let the params select any GLSL version, but that
would entail making glsl::Version implement Deserialize and all that, which
seems like overkill.

Use `back::glsl::features::FeaturesManager` to detect which qualifiers the given
module actually uses, and check them against the requested GLSL version.

The `linear` and `sample` qualifiers for fragment shader inputs are not
supported in all versions of GLSL that Naga can generate.

Non-exhaustive matches on `Expression` make it harder to review changes that add
new enum variants, because the compiler can't tell you that you've missed a
case.

Although I believe Clippy is able to do so, I don't think we should make a
blanket rule of forbidding wildcard match arms against `Expression`. There are a
few matches that are only concerned with a handful of variants, or a single
variant. For example, variant, naga::front::spv::Parser::next_block is only
concerned with whether an expression is a constant or not. In principle, these
sites should also be checked when adding a new variant, but it seems too
burdensome to require them to list all the `Expression` variants they don't care
about. The principle should be applied only to matches that are already handling
most variants.

At the moment, there is only one such match on `Expression` in the code base,
and it's only missing a few variants, so this is quick to fix.

The snapshot files in tests/out start with headers like this:

    ---
    source: tests/snapshots.rs
    expression: dis
    ---
    ; SPIR-V
    ; Version: 1.0
    ; Generator: rspirv
    ...

The Makefile's `validate-foo` targets trim off those lines with `tail -n +5`
before submitting their test to the validators.

However, some versions of the `insta` crate seem to generate headers with an
extra blank line, like this:

    ---
    source: tests/snapshots.rs
    expression: dis

    ---

This makes the `validate-foo` targets unhappy.

This commit changes the `validate-foo` targets to use `sed` to cope with both
kinds of headers.

* Make constant composites non-constexpr

* Update shadow test

* Add a note about references and pointers to Expression::Load

* Formatting

* Remove comma

* Clearer English

* Handle length(float x)

* Handle length(float) and distance(float, float)

* Apply suggestions

* Apply suggestions

* Move stack size range up by 1000

* Take references in function calls if thats what the funciton wants

* Use references instead of pointers the function argument list

* Add hlsl-out feature

* Integrate hlsl-out with snapshots

* Add snapshot

* Handle NaN and inf

* Handle negative infinity

beautify

beautify #2

This adds `use` declarations for `spirv::Decoration` and `spirv::BuiltIn` local
to functions where those are used repeatedly.

The initial intention was to make the code a little more elegant and match on
the Vec itself instead of its length, however there's also a - probably
minuscule - performance advantage:

`Vec::remove` internally copies the removed element onto the stack. In this case
that's a `StructMember` weighting 48 bytes. The `Binding` we are actually
interested in is only 8 bytes tough.

Cloning just the `Binding` saves us from copying 40 bytes for nothing.