[docs] Update pytorch article figure and caption

[ryan-williams]

Correct "default" marker annotation at https://chanzuckerberg.github.io/cellxgene-census/articles/2024/20240709-pytorch.html#increased-training-speed:

gif: before / after

Also made a few tweaks to the caption, to reflect what I think are the most important take-aways:

• ≈2.5x speed improvement vs. shuffling with scipy.csr method (red circles > red squares)
• marginal speed improvement, and decreased memory usage, compared to {sparse, unshuffled} benchmarks (i.e. before #1188 and #1169/#1224; basically red-orange "default" circles vs. corresponding blue squares)

Summarizing another way:

• #1188 introduced shuffling, but ≈halved speed (blue squares → red squares)
• #1224 added 2.5x speedup (red squares → red circles)
• Net result: we added shuffling (⇒ better model training), and marginally increased speed

If a user is willing to use 2x the memory, they can go for the larger red-orange circles (2048 chunks @ 128 rows/chunk), and another nearly ≈2x speedup, but that's not explicitly called out, which is probably fine?

[pablo-gar]

Thanks @ryan-williams Would you mind adding a line right under

*Published:* *July 11th, 2024*

To make it this:

*Published:* *July 11th, 2024*

*Updated:* *July 19th, 2024*. Figure 3 has been improved for readability.