Views with reshaped range indices are linear and strided

[mbauman]

This came up in a DiffEq task, wherein a matrix is stored as a subset of the parameters vector — and thus to be used, it must be reshaped/subsetted before doing a matmul. This patch gives such a view access to the blassy goodness. The exemplar case looks something like this:

p = rand(300)
M = view(p, reshape(2:241, 60, 4))
v = rand(4)
@benchmark $M*$v

Before, this took nearly 200ns. With this patch we hit BLAS and are down to 100ns. Interestingly, this introduces a bifurcation where it's possible for view to be linear but not strided... but fortunately we've been pretty strict on keeping those two attributes distinct even though they previously were 100% overlapping.

I made sure the SubArray tests passed with JULIA_TESTFULL and while I was there I testsetified the test/subarray.jl file (as a separate and independent commit).

cc: @ChrisRackauckas.

[timholy]

This looks great. TBH I only looked carefully at the first commit, I assume that has all the functional changes? Thanks also for the cleanup of the subarray tests, I noticed several nice improvements there.

[ChrisRackauckas]

bump?

[ChrisRackauckas]

Bump again? I'd like to see this in v1.5 if possible.

[ViralBShah]

@mbauman Can we get this in? Just have 4 more days!