support building with clang

[zhanyong-wan]

🚀 Feature

Support building PyTorch/XLA with clang as the C/C++ compiler.

Motivation

Currently the PyTorch/XLA build uses gcc as the C/C++ compiler. Switching to clang gives us several benefits:

• Consistency with PyTorch, which uses clang.
• Faster build - experience shows that clang often is 2-3x faster than gcc.
• Smaller memory usage by the compiler - clang typically uses ~1/3 of the RAM as gcc when compiling.
• Consistency with other tooling we use, like clang-format.

Pitch

1. Add a configuration for building with clang, so that we can add --config=clang to the bazel commandline to select clang.
2. Fix incompatibilities with clang.
3. Measure the compilation performance of clang vs gcc.
4. If the result of 3 is favorable, switch the build to use clang by default.

Alternatives

Do nothing - support gcc only.

[ysiraichi]

This will be actually the case after #8665.

[zhanyong-wan]

Thanks, @ysiraichi ! I think it'd be good to make incremental progress.

[zhanyong-wan]

@bfolie reported that the build is much slower after we switch the default bazel spawn strategy to sandboxed, which is needed by clang. I'll revert the change for now.