Vectorise indexing of contiguous PyArrays

[Beforerr]

#715 was closed because the type parameters can't say an array is contiguous. A runtime check turns out to be enough: branching on strides[1] == sizeof(R) lets LLVM version the loop and vectorise the contiguous case.

Float32, Julia 1.13.1, Apple M1, ms:

	before	after	Array
sum, 1e7 vector	4.43	0.45	0.45
sum, 4-D F-order (4e7)	17.8	2.02	2.02
sum, v[::2]	4.43	4.6

Strided and C-order arrays are unchanged within noise.

Broadcasting into an Array (x .* 2) is still ~3x slower than with an Array input. That one is on the Julia side: LLVM won't emit a runtime alias check between the Array store and the Ptr load.

The new test uses _testbuffer, so it doesn't need numpy.