improve erf performance

[juliantaylor]

as suggested by @pv allow inlining polynomials, combined with unrolling and isnan replace this bumps performance quite considerably:

import numpy as np
import scipy.special as sps
r=np.random.rand(int(1e5))
%timeit -n 1000 -r 5 sps.erf(r)
%timeit -n 1000 -r 5 sps.erf(r + 2.)

before:

1000 loops, best of 5: 2.7 ms per loop
1000 loops, best of 5: 9.56 ms per loop

after

1000 loops, best of 5: 1.17 ms per loop
1000 loops, best of 5: 6.42 ms per loop

[juliantaylor]

this might be worthwhile in many more places than just ndtr, might be a candidate to put into the main header

[pv]

Thanks, merged. Also moved the unroll+builtin-isnan part to apply to whole Cephes.
Didn't seem to cause slowdowns, and at least gamma function was sped up by 30%.

[rgommers]

I ran into this when cleaning up _c99compat.h (see gh-16800), and tested whether this still helps - it indeed does, so I didn't touch it:

In [2]: import numpy as np
   ...: import scipy.special as sps
   ...: r=np.random.rand(int(1e5))
   ...: %timeit -n 1000 -r 5 sps.erf(r)
   ...: %timeit -n 1000 -r 5 sps.erf(r + 2.)
395 µs ± 20.3 µs per loop (mean ± std. dev. of 5 runs, 1,000 loops each)
1.74 ms ± 31.1 µs per loop (mean ± std. dev. of 5 runs, 1,000 loops each)

remove the unroll pragma:

517 µs ± 7.35 µs per loop (mean ± std. dev. of 5 runs, 1,000 loops each)
1.87 ms ± 25.7 µs per loop (mean ± std. dev. of 5 runs, 1,000 loops each)

don't use __builtins_isnan at all:

723 µs ± 4.76 µs per loop (mean ± std. dev. of 5 runs, 1,000 loops each)
2.04 ms ± 7.47 µs per loop (mean ± std. dev. of 5 runs, 1,000 loops each)