Make np.dot and np.inner fast for read-only arrays.

[hansres]

Currently np.dot/np.inner are much slower (~3x) for read-only arrays than for writeable arrays.

In [1]: a = randn(1e6)

In [2]: %timeit dot(a, a)
100 loops, best of 3: 5.42 ms per loop

In [3]: a.setflags(write=False)

In [4]: %timeit dot(a, a)
100 loops, best of 3: 15.9 ms per loop.

This slowdown for read-only inputs is caused by the creation of array copies: dot/inner call PyArray_FromAny and require the WRITABLE flag. Instead, because read-only input arrays are fine, the WRITEABLE flag should not be is not required; and the ALIGNED flag should be sufficient.

[hansres]

Don’t require the WRITEABLE flag (change BEHAVED == ALIGNED|WRITEABLE --> ALIGNED).

[cdavid]

Fixed in r8082