I think @mblesac is right to suggest you should use the b = 1000 s/mm² shell for the FA calculation – that would make it more consistent with the rest of the literature, which is typically single-shell.
It’s also correct that the tensor fit can struggle a bit with multi-shell data, since the signal decay is non-mono-exponential with respect to b (this is after all the entire basis for diffusion kurtosis imaging). How much it struggles depends on the particulars of the acquisition, and crucially on the tensor fitting procedure (as touched on in this post).
These negative values are not uncommon – see this post for an explanation for how these can arise. But these are indeed more likely to occur with higher noise, which matches your experience. Gibbs ringing can also be a major contributor, particularly near the ventricles, due to the much larger CSF signal bringing the surrounding b=0 signal down. You could try using mrdegibbs in addition to dwidenoise, see if that helps further.
As to whether you need to worry about them, personally I don’t think so: I’d recommend you compute FA from the b = 1000 s/mm² shell anyway, so the choice of b-value for the higher shell shouldn’t make much of a difference (other than a minor increase in TE). Whether it might cause problems for your analysis is entirely dependent on what you plan on doing with this (I would personally not recommend doing much with FA values in the first place, but that’s just my opinion… 
).
As to which b-value is best, this is actually a difficult to question to answer with any authority, as things stand. I would personally encourage you to use the higher b-value of 2,500 s/mm², since that has been shown to provide better contrast in the angular domain (helps with estimating fibre orientations). I would expect it probably also provides better contrast in the b-value domain, since that’s often the value used for e.g. diffusion kurtosis imaging, and is also close to the values we picked in the developing human connectome project. There’s also the arguments presented in the FBA papers, although these relate to the interpretation of the results, rather than sensitivity as such – thankfully the recommendations are broadly in line anyway.
Hope this helps answer your questions…?