Fix compiletest's --target-arch handling.

[nnethercote]

compiletests lets you specify one or more target architectures. E.g. on CI we run this:

cargo run -p compiletests --release --no-default-features -- \
    --target-arch compute_61,compute_70,compute_90

While trying out various different target architectures, including invalid ones, I get strange results.

First, if the target arch has fewer than 8 chars (e.g. --target-arch blah), then nvvm panics:

thread 'rustc' panicked at crates/nvvm/src/lib.rs:246:38: byte
index 8 is out of bounds of `blah`

Second, if the target arch has 8 or more chars but is invalid, it is just ignored. This means invalid flags like --target-arch compute_999 run, but just use the default target architecture.

This commit fixes the problems.

• Fix the slicing into the -arch= option to avoid the bounds panic.
• In CodegenArgs::parse, check for the Err case and panic instead of swallowing it. A panic isn't ideal, but it's much better than silently accepting and ignoring invalid input.
• Use a String instead of &'static str for the Err case, so the error message is more informative, i.e. it includes the bad flag value.

Also, the commit improves the NvvmOption::from_str test.

• Move inputs next to expected outputs, which makes it easier to read and update.
• Add the missing compute capabilities.
• Add some checks for invalid inputs.

[LegNeato]

Thanks for catching these