Compiletest normalization

[Firestar99]

Requires #380

To reduce compiletest stderr merge conflicts and reviewing effort:

• Removes line and column from rustc errors:
  

  rust-gpu/tests/compiletests/README.md

  Lines 86 to 87 in bbb1919

  	* `// normalize-stderr-test "\.rs:\d+:\d+" -> ".rs:""` remove the line and column from error messages, eg. `--> path/to/some.rs:8:6`
  	* `// normalize-stderr-test "(\n)\d* *([ -])([\|\+\-\=])" -> "$1 $2$3""` remove the line number prefixing the source code snippet, does not touch `...`

• Removes OpLine from some tests

[LegNeato]

Doesn't appear to be a way we can register this as a default with compiletest and just have a directive like include-file-numbers to turn them on when needed. But there is -Z ui-testing, perhaps we can set that in compiletests's Config https://github.com/Manishearth/compiletest-rs/blob/02b122626320f562914710bada39e637f051c7bb/src/common.rs#L172? do we ever care about the line numbers?

It looks like his might be possible / better in the newer rustc compiletest framework, but I had other issues with it: https://rustc-dev-guide.rust-lang.org/tests/ui.html

[Firestar99]

-Z ui-testing is a good find! Pushed that alternative to compiletest-z-ui-testing, if you want to have a look. The main difference is that the rustc flag keeps the line numbers on the actual error, and just removes the line numbers on the code snippets. I don't really have a preference for which one to merge, feel free to decide.

@eddyb @nnethercote @FractalFir preferences?

(Unfortunately github compare between the branches doesn't actually produce a proper diff, but just shows the diff to the parent commit... so here's a manual sample diff from invalid_vector_type:)

 error: `#[spirv(vector)]` must have 2, 3 or 4 members
-   --> $DIR/invalid_vector_type.rs:
-    |
-    | pub struct FewerFields {
-    | ^^^^^^^^^^^^^^^^^^^^^^
+  --> $DIR/invalid_vector_type.rs:8:1
+   |
+LL | pub struct FewerFields {
+   | ^^^^^^^^^^^^^^^^^^^^^^
 
 error: `#[spirv(vector)]` must have 2, 3 or 4 members
-   --> $DIR/invalid_vector_type.rs:
-    |
-    | pub struct TooManyFields {
-    | ^^^^^^^^^^^^^^^^^^^^^^^^
+  --> $DIR/invalid_vector_type.rs:13:1
+   |
+LL | pub struct TooManyFields {
+   | ^^^^^^^^^^^^^^^^^^^^^^^^
 
 error: `#[spirv(vector)]` type fields must all be floats, integers or bools
-   --> $DIR/invalid_vector_type.rs:
-    |
-    | pub struct NotVectorField {
-    | ^^^^^^^^^^^^^^^^^^^^^^^^^
-    |
-    = note: field type is f32x2
+  --> $DIR/invalid_vector_type.rs:22:1
+   |
+LL | pub struct NotVectorField {
+   | ^^^^^^^^^^^^^^^^^^^^^^^^^
+   |
+   = note: field type is f32x2
 
 error: `#[spirv(vector)]` member types must all be the same
-   --> $DIR/invalid_vector_type.rs:
-    |
-    | pub struct DifferentTypes {
-    | ^^^^^^^^^^^^^^^^^^^^^^^^^
+  --> $DIR/invalid_vector_type.rs:34:1
+   |
+LL | pub struct DifferentTypes {
+   | ^^^^^^^^^^^^^^^^^^^^^^^^^
 
 error: aborting due to 4 previous errors
 

We may need to modernize our compiletest at some point, but I don't think we should block this PR because of that. I'm mostly waiting for my difftest PR to be reviewed before I'll touch it, so I can immediately integrate the feedback from there.

[nnethercote]

I'm used to the LL style from rustc, I like that.

Also, I hit some of this stuff in the Nightly update in #420, so good to see it'll stop being a source of churn.

[Firestar99]

Then let's do rustc's LL style.

Closed in favor of #447