Split up long running tests some more

We have some rather long-running tests, particularly on slow machines (e.g. KNL). It would be good to split those up some more. The worst offenders (from a KNL run):

linalg/diagonal (14)        |  510.55  |  1.72  |  0.3 | 1700.32    | 338.29
bitarray (76)               |  526.51  |  2.81  |  0.5 | 2722.84    | 394.37
linalg/qr (3)               |  570.19  |  3.46  |  0.6 | 3087.33    | 377.35
arrayops (66)               |  569.20  |  3.80  |  0.7 | 2473.68    | 485.29
linalg/lq (22)              |  576.79  |  1.86  |  0.3 | 3036.09    | 378.71
parallel (146)              |  669.16  |  0.00  |  0.0 |  5.18      | 237.79
sparse/sparse (27)          |  711.90  | 70.70  |  9.9 | 2044.43    | 418.86
linalg/dense (4)            |  801.39  |  3.59  |  0.4 | 3180.17    | 388.57
sparse/higherorderfns (29)  |  804.48  |  3.51  |  0.4 | 3512.64    | 434.74
sparse/sparsevector (28)    |  881.77  |  2.81  |  0.3 | 2846.67    | 389.55
linalg/arnoldi (25)         |  991.20  |  1.59  |  0.2 | 1185.46    | 318.51
linalg/matmul (5)           | 1403.10  |  1.89  |  0.1 | 2363.30    | 357.77
subarray (26)               | 1450.06  | 16.15  |  1.1 | 8445.22    | 633.14
linalg/triangular (2)       | 2040.71  | 28.97  |  1.4 | 10229.32   | 603.94

Would be good to split up everything over 500s or so. For extra credit, keep these in the same file, but have some mechanism to declare a split.