Update to GCC 11

[bdice]

This PR updates builds to use GCC 11.

[bdice]

https://github.com/rapidsai/cuml/actions/runs/4297821777/jobs/7491310331#step:6:1107

Error log:

In function 'void UMAPAlgo::Optimize::optimize_params(T*, int, const T*, T*, ML::UMAPParams*, cudaStream_t, float, int) [with T = float; int TPB_X = 256]',
    inlined from 'void UMAPAlgo::Optimize::find_params_ab(ML::UMAPParams*, cudaStream_t)' at $SRC_DIR/cpp/src/umap/optimize.cuh:208:30:
$SRC_DIR/cpp/src/umap/optimize.cuh:167:1: error: 'void operator delete(void*, std::size_t)' called on pointer returned from a mismatched allocation function [-Werror=mismatched-new-delete]
  167 |     delete grads_h;
      | ^   ~~~~~~~~~~
$SRC_DIR/cpp/src/umap/optimize.cuh: In function 'void UMAPAlgo::Optimize::find_params_ab(ML::UMAPParams*, cudaStream_t)':
$SRC_DIR/cpp/src/umap/optimize.cuh:156:25: note: returned from 'void* malloc(size_t)'
  156 |     T* grads_h = (T*)malloc(2 * sizeof(T));
      |                   ~~~~~~^~~~~~~~~~~~~~~~~
cc1plus: all warnings being treated as errors

[bdice]

Waiting on #5259, then we can update this and CI should pass.

This is ready for review, but please do not merge until all RAPIDS repos are ready to migrate. (I do not have privileges to add the DO NOT MERGE label to prohibit merging.)

[raydouglass]

/merge

[wence-]

This didn't merge due to a single failure test_concat_memory_leak[large_clf0-regression]:
assert (5631447040 - 5630435328) < 1000000.0 which is false because 5631447040 - 5630435328 == 1011712. Is this a tolerance issue one can relax?

[bdice]

@wence- Apparently my last commit passed that threshold, triggering the merge. That seems like it might be a flaky test?