Hm, I think the understanding that "$5 spent to buy a net has marginal impact, but $5 to an EA fund does not" isn't quite right. You may be correct that the $5 has a lower probability of changing anything at the margins in a fund, but you should multiply that probability by the expected impact of the change.

To pick an analogy from the animal suffering wing: is it better to reduce $5 spent on a chicken or $5 on a cow? Naively you might think "a cow costs $5000, chicken costs $5, so at the margins you should always pick chicken, cuz that's guaranteed to do something". But that excludes the expected value of the impact; any specific cow purchase only has a 1/1000 chance of being the one that pushes it over the line, but if (for example) the cow life was actually worth 2000x more than a chicken, than the expected value of your impact is better on the cow.

(Yes, I'm making lots of simplifying assumptions here - just trying to provide the intuition, not actually do a moral analysis!)

I would guess a marginal donation usually effectively goes to one of the charities they would have funded anyway (in one period or another), not necessarily the first charity they would have not funded if they had less to donate, and sometimes to a charity they wouldn't have funded otherwise, which would be less attractive than the ones they would have funded anyway. You can think in terms of shifting the distributions of charities and payouts if you want to avoid butterfly effects.

You might not expect decreasing marginal returns, in case the charities themselves have increasing marginal returns. You would expect decreasing marginal returns when your donation goes to a marginal charity or if the charities have decreasing marginal returns.

Then it becomes a matter of which charities and what those charities do with marginal donations.

I think the issue of marginal impact, particularly for relatively small donations going to organizations with rather large budget/granting power, is quite difficult to be confident in, and hence shouldn't be the driving factor in looking at the benefit of said donation. Much of the problem is illustrated well in Budolfson's and Spear's essay, The Hidden Zero Problem. 

It does seem that many EA calculation rely on one of Parfit's 'share of the total' errors that he outlined in Reasons and Persons, making transparent advertisement about the effectiveness of individual donations quite difficult.