Fix sccache for CTK 11.1 and properly track compilations in stats

[trxcllnt]

This PR has some fixes I neglected to add to #2247.

1. nvcc in CUDA toolkit v11.1 didn't add the -D__CUDA_ARCH_LIST__= definition, so 5271494 expands the list of defines that indicate an nvcc host compiler invocation.
2. f160a0a and 1591089 report compilation type (local or dist) and duration for forced-no-cache, forced-recache, and compilation failures. It also counts and reports total compilations performed, not just compilations due to cache misses.
3. ccfc60b ensures compilations with --verbose are never dist-compiled, since the verbose output is parsed by tools like CMake and must reflect the local toolchain.
4. bdaf35e adds more clang flags so using clang as a CUDA compiler with -Xclang doesn't fail

Question for @sylvestre related to the last point -- do you know which bits of the clang toolchain (or CTK?) sccache should package when using clang as a device compiler? I am seeing errors like the following when attempting to dist-compile with ClangCUDA, but I'm not sure which files define the __nvvm_* symbols:

In file included from build/libcudacxx/test/internal_headers/headers/__barrier_async_contract_fulfillment.h.cu:1:
In file included from <built-in>:1:
In file included from /usr/lib/llvm-18/lib/clang/18/include/__clang_cuda_runtime_wrapper.h:73:
/usr/lib/llvm-18/lib/clang/18/include/__clang_cuda_builtin_vars.h:53:180: error: use of undeclared identifier '__nvvm_read_ptx_sreg_tid_x'
   53 |   __declspec(property(get = __fetch_builtin_x)) unsigned int x; static inline __attribute__((always_inline)) __attribute__((device)) unsigned int __fetch_builtin_x(void) { return __nvvm_read_ptx_sreg_tid_x(); };
...

[sylvestre]

sorry, i don't know

[trxcllnt]

This doesn't seem to be an issue with sccache. It appears clang can't compile its own preprocessor output:

#!/usr/bin/env bash

# Basic CUDA example from https://godbolt.org/
cat <<EOF >/tmp/test.cu
__global__ void square(int* array, int n) {
    int tid = blockDim.x * blockIdx.x + threadIdx.x;
    if (tid < n)
        array[tid] = array[tid] * array[tid];
}
EOF

# Preprocess
clang++ -x cuda -E --cuda-gpu-arch=sm_80 --cuda-path=/usr/local/cuda -Wno-unknown-cuda-version /tmp/test.cu > /tmp/test.cui

# Compile (fails)
clang++ -x cuda-cpp-output --cuda-gpu-arch=sm_80 --cuda-path=/usr/local/cuda -Wno-unknown-cuda-version -o /tmp/test.cu.o /tmp/test.cui

[trxcllnt]

cc: @robertmaynard for review

[trxcllnt]

This PR is ready to merge.

[sylvestre]

Given that the initial implementation had issues and you are fixing them here, could you please add tests to cover these missing cases?
thanks

[trxcllnt]

@sylvestre I added tests to ensure -v|--verbose and clang-cuda compilations aren't dist-compiled, and added CUDA Toolkit v11.1 to the CI jobs to ensure we accommodate nvcc changes introduced between CTK v11.1 and v11.8.

CTK v11.1 is the oldest version we test in CUDA Core Compute Libraries. I wouldn't mind testing earlier versions, but CTK v10.2 is only available up to ubuntu18.04, GH has removed ubuntu18.04 runners, so we'd have to containerize the test jobs.