correct handling of KNOWN_SOLVER_ERRORS in test_constant_atoms.py

[rileyjmurray]

The build failures in https://github.com/cvxpy/cvxpy/actions/runs/3772470280/jobs/6413421816 happened even though (cp.xexp, cp.MOSEK) were added to the list of known solver errors in test_constant_atoms.py. Upon further inspection, that's because the KNOWN_SOLVER_ERRORS list wasn't being checked in the appropriate way. This PR addresses that; I've confirmed that it correctly skips the case of ('xexp', cp.MOSEK) (note the problematic atom has to be specified as a string now).

[rileyjmurray]

This should resolve #1983.

[github-actions]

Benchmarks that have stayed the same:

   before           after         ratio
 [3310da69]       [33b89827]
      1.98±0s          2.07±0s     1.04  simple_LP_benchmarks.SimpleScalarParametrizedLPBenchmark.time_compile_problem
      13.7±0s          14.1±0s     1.03  svm_l1_regularization.SVMWithL1Regularization.time_compile_problem
      10.3±0s          10.5±0s     1.02  semidefinite_programming.SemidefiniteProgramming.time_compile_problem
      56.9±0s          57.8±0s     1.02  sdp_segfault_1132_benchmark.SDPSegfault1132Benchmark.time_compile_problem
      4.65±0s          4.69±0s     1.01  simple_LP_benchmarks.SimpleFullyParametrizedLPBenchmark.time_compile_problem
      26.3±0s          26.5±0s     1.01  qp_1611_benchmark.QP1611Benchmark.time_compile_problem
      35.9±0s          36.1±0s     1.01  cvar_benchmark.CVaRBenchmark.time_compile_problem
      27.9±0s          28.0±0s     1.00  simple_LP_benchmarks.SimpleLPBenchmark.time_compile_problem
      8.44±0s          8.45±0s     1.00  optimal_advertising.OptimalAdvertising.time_compile_problem
      12.6±0s          12.5±0s     0.99  huber_regression.HuberRegression.time_compile_problem