Speed up blurring_mask_2d_from for large PSF kernels

[GeorgeVassilakis]

Hi James!

I'm George, a PhD student at Cambridge. I've been using PyAutoLens for my HWO-SLAPS project on subhalo detectability with the Habitable Worlds Observatory (paper here). I worked with your former collaborators Richard Massey and Qiuhan He on this :)

The project uses really large PSF kernels (up to 999 x 999), and building the convolver got extremely slow and memory hungry as the kernel grew, to the point of OOM-killing jobs on our cluster. I traced it to the binary_dilation in blurring_mask_2d_from, which uses a dense np.ones((ky, kx)) structure (scipy's offsets table for it grows as (ky * kx)**2, which is where the memory goes). Since the structure is a box, the dilation is separable, so a (ky, 1) dilation followed by a (1, kx) dilation gives the exact same mask for a fraction of the cost:

kernel	time before	time after	memory before	memory after
11 x 11	0.021 s	0.020 s	2.2 MiB	2.2 MiB
101 x 101	0.99 s	0.032 s	794 MiB	2.7 MiB
201 x 201	13.2 s	0.088 s	12.2 GiB	3.6 MiB
301 x 301	66.5 s	0.197 s	61.2 GiB	4.9 MiB
499 x 499	401 s	0.67 s	462 GiB	8.7 MiB
999 x 999	not run	4.49 s	~7.2 TiB (estimated)	25.5 MiB

The masks are byte-identical, the full test_autoarray suite passes, and I added a test that checks against the dense dilation.

For the paper I just patched autoarray locally, but I figured it'd be better to get this upstream so anyone else using big PSFs doesn't hit the same wall. I have a PR ready to go if you're happy with the approach. Thanks for all the work on PyAuto, and let me know what you think!