Only benchmark on PRs

[ericphanson]

This prevents also benchmarking again after merging a PR, which is slow and prevents Travis from running other jobs.

[ericphanson]

Benchmarks (which should all be invariant):

               ["ECOS", "GT_constraints"] 0.65 (5%) ✅    1.00 (1%)
               ["ECOS", "LT_constraints"] 0.70 (5%) ✅    1.00 (1%)
             ["ECOS", "affine_diag_atom"] 0.92 (5%) ✅    1.00 (1%)
            ["ECOS", "affine_index_atom"] 0.94 (5%) ✅    1.00 (1%)
              ["ECOS", "affine_sum_atom"] 0.94 (5%) ✅    1.00 (1%)
               ["ECOS", "affine_tr_atom"] 0.90 (5%) ✅    1.00 (1%)
         ["ECOS", "equality_constraints"] 0.67 (5%) ✅    1.00 (1%)
                ["SCS", "GT_constraints"] 0.73 (5%) ✅    1.00 (1%)
                 ["SCS", "LT_constraint"] 0.01 (5%) ✅  0.00 (1%) ✅
                ["SCS", "LT_constraints"] 0.76 (5%) ✅    1.00 (1%)
               ["SCS", "SDP_constraints"] 0.78 (5%) ✅    1.00 (1%)
               ["SCS", "affine_dot_atom"] 0.91 (5%) ✅    1.00 (1%)
             ["SCS", "affine_index_atom"] 0.94 (5%) ✅    1.00 (1%)
              ["SCS", "affine_mult_atom"] 0.91 (5%) ✅    1.00 (1%)
                ["SCS", "affine_tr_atom"] 0.94 (5%) ✅    1.00 (1%)
         ["SCS", "affine_transpose_atom"] 0.94 (5%) ✅    1.00 (1%)
          ["SCS", "equality_constraints"] 0.73 (5%) ✅    1.00 (1%)
           ["SCS", "sdp_lambda_max_atom"] 0.95 (5%) ✅    1.00 (1%)
           ["SCS", "sdp_lambda_min_atom"] 0.94 (5%) ✅    1.00 (1%)
        ["formulation", "GT_constraints"] 0.66 (5%) ✅    1.00 (1%)
        ["formulation", "LT_constraints"] 0.64 (5%) ✅    1.00 (1%)
       ["formulation", "affine_add_atom"] 0.95 (5%) ✅    1.00 (1%)
      ["formulation", "affine_diag_atom"] 0.95 (5%) ✅    1.00 (1%)
      ["formulation", "affine_kron_atom"] 0.94 (5%) ✅    1.00 (1%)
        ["formulation", "affine_tr_atom"] 0.93 (5%) ✅    1.00 (1%)
  ["formulation", "equality_constraints"] 0.66 (5%) ✅    1.00 (1%)
   ["formulation", "sdp_lambda_min_atom"] 0.94 (5%) ✅    1.00 (1%)
["formulation", "sdp_operator_norm_atom"] 0.91 (5%) ✅    1.00 (1%)

Looks worse than #323; again the *constraints benchmarks show up, but now so do several atoms too.