Use stdlib-isystem for cxx stlib include path

[cerisier]

This change needs careful review due to its impact on correctness, long-term maintenance, and downstream users of the toolchain (e.g. rules_go, rules_foreign_cc).

Background

Until now, we passed C++ include paths via -isystem. This worked because Clang implicitly appends its builtin headers at the end of the search path (and we also pass them explicitly via -internal-isystem).

In CUDA mode (-x cuda), the driver injects an additional builtin include path (cuda_wrappers) before the standard C++ includes. This ordering is required because these wrappers override parts of the C++ standard library specifically for CUDA.

With -isystem, we lose control over this ordering, which breaks CUDA compilation.

The problem

We need to preserve Clang’s intended include ordering in CUDA mode:

<-isystem paths>
<clang/.../cuda_wrappers>
<-stdlib++-isystem paths>
<clang/.../include>

Options considered

1. Manually inject cuda_wrappers

This would restore the correct ordering, but requires reliably detecting CUDA mode.

That is non-trivial:

• A cuda_compile action is being introduced, but
• Many projects (e.g. XLA) compile .cc files with -x cuda, which bypasses extension-based detection

This is a good option but it means we can never be compliant with such codebases unless they rename their files to .cu or .cu.cc.

2. Rely on the Clang driver (-stdlib++-isystem) (chosen)

-stdlib++-isystem lets the driver handle ordering of standard library includes correctly, including CUDA-specific adjustments.

However, this introduces a constraint:

• -stdlib++-isystem is ignored in C mode

What this PR does

• In C++ (cpp_compile):

  • Pass all relevant include paths (stdlib++, libc, kernel) via -stdlib++-isystem
  • This ensures correct ordering relative to cuda_wrappers

• In C (c_compile):

  • Continue using -isystem for libc and kernel
  • Since -stdlib++-isystem is ignored

Tradeoff

We are left with two imperfect options:

1. Detect CUDA and manually inject cuda_wrappers

   • Requires reliable CUDA mode detection → works but not in all cases.

2. Delegate ordering to the driver via -stdlib++-isystem

   • Correct behavior
   • But forces divergence between C and C++ include handling

This PR adopts option (2) as the only robust approach given current constraints.

[cerisier]

Also PR fails because windows mingw headers need also be passed -stdlib++-isystem...